Anirudha Majumdar

Amir Ali Ahmadi, Anirudha Majumdar, Russ Tedrake · mit

We show that for continuous time dynamical systems described by polynomial differential equations of modest degree (typically equal to three), the following decision problems which arise in numerous areas of systems and control theory cannot have a polynomial time (or even pseudo-polynomial time) algorithm unless P=NP: local attractivity of an equilibrium point, stability of an equilibrium point in the sense of Lyapunov, boundedness of trajectories, convergence of all trajectories in a ball to a given equilibrium point, existence of a quadratic Lyapunov function, invariance of a ball, invariance of a quartic semialgebraic set under linear dynamics, local collision avoidance, and existence of a stabilizing control law. We also extend our earlier NP-hardness proof of testing local asymptotic stability for polynomial vector fields to the case of trigonometric differential equations of degree four.

Allen Z. Ren, Anushri Dixit, Alexandra Bodrova et al.

Large language models (LLMs) exhibit a wide range of promising capabilities -- from step-by-step planning to commonsense reasoning -- that may provide utility for robots, but remain prone to confidently hallucinated predictions. In this work, we present KnowNo, which is a framework for measuring and aligning the uncertainty of LLM-based planners such that they know when they don't know and ask for help when needed. KnowNo builds on the theory of conformal prediction to provide statistical guarantees on task completion while minimizing human help in complex multi-step planning settings. Experiments across a variety of simulated and real robot setups that involve tasks with different modes of ambiguity (e.g., from spatial to numeric uncertainties, from human preferences to Winograd schemas) show that KnowNo performs favorably over modern baselines (which may involve ensembles or extensive prompt tuning) in terms of improving efficiency and autonomy, while providing formal assurances. KnowNo can be used with LLMs out of the box without model-finetuning, and suggests a promising lightweight approach to modeling uncertainty that can complement and scale with the growing capabilities of foundation models. Website: https://robot-help.github.io

Jensen Gao, Bidipta Sarkar, Fei Xia et al.

Recent advances in vision-language models (VLMs) have led to improved performance on tasks such as visual question answering and image captioning. Consequently, these models are now well-positioned to reason about the physical world, particularly within domains such as robotic manipulation. However, current VLMs are limited in their understanding of the physical concepts (e.g., material, fragility) of common objects, which restricts their usefulness for robotic manipulation tasks that involve interaction and physical reasoning about such objects. To address this limitation, we propose PhysObjects, an object-centric dataset of 39.6K crowd-sourced and 417K automated physical concept annotations of common household objects. We demonstrate that fine-tuning a VLM on PhysObjects improves its understanding of physical object concepts, including generalization to held-out concepts, by capturing human priors of these concepts from visual appearance. We incorporate this physically grounded VLM in an interactive framework with a large language model-based robotic planner, and show improved planning performance on tasks that require reasoning about physical object concepts, compared to baselines that do not leverage physically grounded VLMs. We additionally illustrate the benefits of our physically grounded VLM on a real robot, where it improves task success rates. We release our dataset and provide further details and visualizations of our results at https://iliad.stanford.edu/pg-vlm/.

Y. Isabel Liu, Windsor Nguyen, Yagiz Devre et al. · princeton

Recent advances in state-space model architectures have shown great promise for efficient sequence modeling, but challenges remain in balancing computational efficiency with model expressiveness. We propose the Flash STU architecture, a hybrid model that interleaves spectral state space model layers with sliding window attention, enabling scalability to billions of parameters for language modeling while maintaining a near-linear time complexity. We evaluate the Flash STU and its variants on diverse sequence prediction tasks, including linear dynamical systems, robotics control, and language modeling. We find that, given a fixed parameter budget, the Flash STU architecture consistently outperforms the Transformer and other leading state-space models such as S4 and Mamba-2.

Huiqiong Li, Jiayu Wang, Zhiting Mei et al.

Video world models are increasingly used in robotic manipulation, yet existing benchmarks mostly evaluate them under valid, feasible, and safe instructions. We introduce RoboTrustBench, a benchmark for evaluating the trustworthiness of video world models under four scenarios: Normal, Constraint-Sensitive, Counterfactual, and Adversarial. Built from real-world DROID episodes, RoboTrustBench contains 1,207 expert-validated instruction-image pairs and a six-dimensional evaluation protocol with 13 fine-grained criteria. Evaluating seven representative video world models with human and MLLM assessment, we find that current models often generate visually coherent videos, but struggle with constraint reasoning, counterfactual grounding, physical interaction, and unsafe-instruction suppression. These results show that visual quality and surface-level instruction following are insufficient for trustworthy robotic video world modeling.

Sumeet Singh, Yin-Lam Chow, Anirudha Majumdar et al.

In this paper we present a framework for risk-sensitive model predictive control (MPC) of linear systems affected by stochastic multiplicative uncertainty. Our key innovation is to consider a time-consistent, dynamic risk evaluation of the cumulative cost as the objective function to be minimized. This framework is axiomatically justified in terms of time-consistency of risk assessments, is amenable to dynamic optimization, and is unifying in the sense that it captures a full range of risk preferences from risk-neutral (i.e., expectation) to worst case. Within this framework, we propose and analyze an online risk-sensitive MPC algorithm that is provably stabilizing. Furthermore, by exploiting the dual representation of time-consistent, dynamic risk measures, we cast the computation of the MPC control law as a convex optimization problem amenable to real-time implementation. Simulation results are presented and discussed.

Allen Z. Ren, Bharat Govil, Tsung-Yen Yang et al.

Robust and generalized tool manipulation requires an understanding of the properties and affordances of different tools. We investigate whether linguistic information about a tool (e.g., its geometry, common uses) can help control policies adapt faster to new tools for a given task. We obtain diverse descriptions of various tools in natural language and use pre-trained language models to generate their feature representations. We then perform language-conditioned meta-learning to learn policies that can efficiently adapt to new tools given their corresponding text descriptions. Our results demonstrate that combining linguistic information and meta-learning significantly accelerates tool learning in several manipulation tasks including pushing, lifting, sweeping, and hammering.

Allen Z. Ren, Justin Lidard, Lars L. Ankile et al.

We introduce Diffusion Policy Policy Optimization, DPPO, an algorithmic framework including best practices for fine-tuning diffusion-based policies (e.g. Diffusion Policy) in continuous control and robot learning tasks using the policy gradient (PG) method from reinforcement learning (RL). PG methods are ubiquitous in training RL policies with other policy parameterizations; nevertheless, they had been conjectured to be less efficient for diffusion-based policies. Surprisingly, we show that DPPO achieves the strongest overall performance and efficiency for fine-tuning in common benchmarks compared to other RL methods for diffusion-based policies and also compared to PG fine-tuning of other policy parameterizations. Through experimental investigation, we find that DPPO takes advantage of unique synergies between RL fine-tuning and the diffusion parameterization, leading to structured and on-manifold exploration, stable training, and strong policy robustness. We further demonstrate the strengths of DPPO in a range of realistic settings, including simulated robotic tasks with pixel observations, and via zero-shot deployment of simulation-trained policies on robot hardware in a long-horizon, multi-stage manipulation task. Website with code: diffusion-ppo.github.io

Allen Z. Ren, Hongkai Dai, Benjamin Burchfiel et al.

Simulation parameter settings such as contact models and object geometry approximations are critical to training robust robotic policies capable of transferring from simulation to real-world deployment. Previous approaches typically handcraft distributions over such parameters (domain randomization), or identify parameters that best match the dynamics of the real environment (system identification). However, there is often an irreducible gap between simulation and reality: attempting to match the dynamics between simulation and reality across all states and tasks may be infeasible and may not lead to policies that perform well in reality for a specific task. Addressing this issue, we propose AdaptSim, a new task-driven adaptation framework for sim-to-real transfer that aims to optimize task performance in target (real) environments -- instead of matching dynamics between simulation and reality. First, we meta-learn an adaptation policy in simulation using reinforcement learning for adjusting the simulation parameter distribution based on the current policy's performance in a target environment. We then perform iterative real-world adaptation by inferring new simulation parameter distributions for policy training, using a small amount of real data. We perform experiments in three robotic tasks: (1) swing-up of linearized double pendulum, (2) dynamic table-top pushing of a bottle, and (3) dynamic scooping of food pieces with a spatula. Our extensive simulation and hardware experiments demonstrate AdaptSim achieving 1-3x asymptotic performance and $\sim$2x real data efficiency when adapting to different environments, compared to methods based on Sys-ID and directly training the task policy in target environments. Website: https://irom-lab.github.io/AdaptSim/

Tenny Yin, Zhiting Mei, Zhonghe Zheng et al.

Action-conditioned video models offer a promising path to building general-purpose robot simulators that can improve directly from data. Yet, despite training on large-scale robot datasets, current state-of-the-art video models still struggle to predict physically consistent robot-object interactions that are crucial in robotic manipulation. To close this gap, we present PlayWorld, a simple, scalable, and fully autonomous pipeline for training high-fidelity video world simulators from interaction experience. In contrast to prior approaches that rely on success-biased human demonstrations, PlayWorld is the first system capable of learning entirely from unsupervised robot self-play, enabling naturally scalable data collection while capturing complex, long-tailed physical interactions essential for modeling realistic object dynamics. Experiments across diverse manipulation tasks show that PlayWorld generates high-quality, physically consistent predictions for contact-rich interactions that are not captured by world models trained on human-collected data. We further demonstrate the versatility of PlayWorld in enabling fine-grained failure prediction and policy evaluation, with up to 40% improvements over human-collected data. Finally, we demonstrate how PlayWorld enables reinforcement learning in the world model, improving policy performance by 65% in success rates when deployed in the real world.

Anirudha Majumdar

We seek to understand fundamental tradeoffs between the accuracy of prior information that a learner has on a given problem and its learning performance. We introduce the notion of prioritized risk, which differs from traditional notions of minimax and Bayes risk by allowing us to study such fundamental tradeoffs in settings where reality does not necessarily conform to the learner's prior. We present a general reduction-based approach for extending classical minimax lower-bound techniques in order to lower bound the prioritized risk for statistical estimation problems. We also introduce a novel generalization of Fano's inequality (which may be of independent interest) for lower bounding the prioritized risk in more general settings involving unbounded losses. We illustrate the ability of our framework to provide insights into tradeoffs between prior information and learning performance for problems in estimation, regression, and reinforcement learning.

David Snyder, Apurva Badithela, Nikolai Matni et al.

Generalist robot manipulation policies are becoming increasingly capable, but are limited in evaluation to a small number of hardware rollouts. This strong resource constraint in real-world testing necessitates both more informative performance measures and reliable and efficient evaluation procedures to properly assess model capabilities and benchmark progress in the field. This work presents a novel framework for robot policy comparison that is sample-efficient, statistically rigorous, and applicable to a broad set of evaluation metrics used in practice. Based on safe, anytime-valid inference (SAVI), our test procedure is sequential, allowing the evaluator to stop early when sufficient statistical evidence has accumulated to reach a decision at a pre-specified level of confidence. Unlike previous work developed for binary success, our unified approach addresses a wide range of informative metrics: from discrete partial credit task progress to continuous measures of episodic reward or trajectory smoothness, spanning both parametric and nonparametric comparison problems. Through extensive validation on simulated and real-world evaluation data, we demonstrate up to 70% reduction in evaluation burden compared to standard batch methods and up to 50% reduction compared to state-of-the-art sequential procedures designed for binary outcomes, with no loss of statistical rigor. Notably, our empirical results show that competing policies can be separated more quickly when using fine-grained task progress than binary success metrics.

Lihan Zha, Asher J. Hancock, Mingtong Zhang et al.

A long-standing goal in robotics is a generalist policy that can be deployed zero-shot on new robot embodiments without per-embodiment adaptation. Despite large-scale multi-embodiment pre-training, existing Vision-Language-Action models (VLAs) remain tightly coupled to their training embodiments and typically require costly fine-tuning. We introduce Language-Action Pre-training (LAP), a simple recipe that represents low-level robot actions directly in natural language, aligning action supervision with the pre-trained vision-language model's input-output distribution. LAP requires no learned tokenizer, no costly annotation, and no embodiment-specific architectural design. Based on LAP, we present LAP-3B, which to the best of our knowledge is the first VLA to achieve substantial zero-shot transfer to previously unseen robot embodiments without any embodiment-specific fine-tuning. Across multiple novel robots and manipulation tasks, LAP-3B attains over 50% average zero-shot success, delivering roughly a 2x improvement over the strongest prior VLAs. We further show that LAP enables efficient adaptation and favorable scaling, while unifying action prediction and VQA in a shared language-action format that yields additional gains through co-training.

Paula Gradu, John Hallman, Daniel Suo et al.

We present an open-source library of natively differentiable physics and robotics environments, accompanied by gradient-based control methods and a benchmark-ing suite. The introduced environments allow auto-differentiation through the simulation dynamics, and thereby permit fast training of controllers. The library features several popular environments, including classical control settings from OpenAI Gym. We also provide a novel differentiable environment, based on deep neural networks, that simulates medical ventilation. We give several use-cases of new scientific results obtained using the library. This includes a medical ventilator simulator and controller, an adaptive control method for time-varying linear dynamical systems, and new gradient-based methods for control of linear dynamical systems with adversarial perturbations.

Allen Z. Ren, Jaden Clark, Anushri Dixit et al.

We consider the problem of Embodied Question Answering (EQA), which refers to settings where an embodied agent such as a robot needs to actively explore an environment to gather information until it is confident about the answer to a question. In this work, we leverage the strong semantic reasoning capabilities of large vision-language models (VLMs) to efficiently explore and answer such questions. However, there are two main challenges when using VLMs in EQA: they do not have an internal memory for mapping the scene to be able to plan how to explore over time, and their confidence can be miscalibrated and can cause the robot to prematurely stop exploration or over-explore. We propose a method that first builds a semantic map of the scene based on depth information and via visual prompting of a VLM - leveraging its vast knowledge of relevant regions of the scene for exploration. Next, we use conformal prediction to calibrate the VLM's question answering confidence, allowing the robot to know when to stop exploration - leading to a more calibrated and efficient exploration strategy. To test our framework in simulation, we also contribute a new EQA dataset with diverse, realistic human-robot scenarios and scenes built upon the Habitat-Matterport 3D Research Dataset (HM3D). Both simulated and real robot experiments show our proposed approach improves the performance and efficiency over baselines that do no leverage VLM for exploration or do not calibrate its confidence. Webpage with experiment videos and code: https://explore-eqa.github.io/

Ola Shorinwa, Zhiting Mei, Justin Lidard et al.

The remarkable performance of large language models (LLMs) in content generation, coding, and common-sense reasoning has spurred widespread integration into many facets of society. However, integration of LLMs raises valid questions on their reliability and trustworthiness, given their propensity to generate hallucinations: plausible, factually-incorrect responses, which are expressed with striking confidence. Previous work has shown that hallucinations and other non-factual responses generated by LLMs can be detected by examining the uncertainty of the LLM in its response to the pertinent prompt, driving significant research efforts devoted to quantifying the uncertainty of LLMs. This survey seeks to provide an extensive review of existing uncertainty quantification methods for LLMs, identifying their salient features, along with their strengths and weaknesses. We present existing methods within a relevant taxonomy, unifying ostensibly disparate methods to aid understanding of the state of the art. Furthermore, we highlight applications of uncertainty quantification methods for LLMs, spanning chatbot and textual applications to embodied artificial intelligence applications in robotics. We conclude with open research challenges in uncertainty quantification of LLMs, seeking to motivate future research.

Apoorva Sharma, Sushant Veer, Asher Hancock et al.

Inductive Conformal Prediction (ICP) provides a practical and effective approach for equipping deep learning models with uncertainty estimates in the form of set-valued predictions which are guaranteed to contain the ground truth with high probability. Despite the appeal of this coverage guarantee, these sets may not be efficient: the size and contents of the prediction sets are not directly controlled, and instead depend on the underlying model and choice of score function. To remedy this, recent work has proposed learning model and score function parameters using data to directly optimize the efficiency of the ICP prediction sets. While appealing, the generalization theory for such an approach is lacking: direct optimization of empirical efficiency may yield prediction sets that are either no longer efficient on test data, or no longer obtain the required coverage on test data. In this work, we use PAC-Bayes theory to obtain generalization bounds on both the coverage and the efficiency of set-valued predictors which can be directly optimized to maximize efficiency while satisfying a desired test coverage. In contrast to prior work, our framework allows us to utilize the entire calibration dataset to learn the parameters of the model and score function, instead of requiring a separate hold-out set for obtaining test-time coverage guarantees. We leverage these theoretical results to provide a practical algorithm for using calibration data to simultaneously fine-tune the parameters of a model and score function while guaranteeing test-time coverage and efficiency of the resulting prediction sets. We evaluate the approach on regression and classification tasks, and outperform baselines calibrated using a Hoeffding bound-based PAC guarantee on ICP, especially in the low-data regime.

Zhiting Mei, Christina Zhang, Tenny Yin et al.

Reasoning language models have set state-of-the-art (SOTA) records on many challenging benchmarks, enabled by multi-step reasoning induced using reinforcement learning. However, like previous language models, reasoning models are prone to generating confident, plausible responses that are incorrect (hallucinations). Knowing when and how much to trust these models is critical to the safe deployment of reasoning models in real-world applications. To this end, we explore uncertainty quantification of reasoning models in this work. Specifically, we ask three fundamental questions: First, are reasoning models well-calibrated? Second, does deeper reasoning improve model calibration? Finally, inspired by humans' innate ability to double-check their thought processes to verify the validity of their answers and their confidence, we ask: can reasoning models improve their calibration by explicitly reasoning about their chain-of-thought traces? We introduce introspective uncertainty quantification (UQ) to explore this direction. In extensive evaluations on SOTA reasoning models across a broad range of benchmarks, we find that reasoning models: (i) are typically overconfident, with self-verbalized confidence estimates often greater than 85% particularly for incorrect responses, (ii) become even more overconfident with deeper reasoning, and (iii) can become better calibrated through introspection (e.g., o3-Mini and DeepSeek R1) but not uniformly (e.g., Claude 3.7 Sonnet becomes more poorly calibrated). Lastly, we conclude with important research directions to design necessary UQ benchmarks and improve the calibration of reasoning models.

Anirudha Majumdar, Mohit Sharma, Dmitry Kalashnikov et al.

Visuomotor policies trained via imitation learning are capable of performing challenging manipulation tasks, but are often extremely brittle to lighting, visual distractors, and object locations. These vulnerabilities can depend unpredictably on the specifics of training, and are challenging to expose without time-consuming and expensive hardware evaluations. We propose the problem of predictive red teaming: discovering vulnerabilities of a policy with respect to environmental factors, and predicting the corresponding performance degradation without hardware evaluations in off-nominal scenarios. In order to achieve this, we develop RoboART: an automated red teaming (ART) pipeline that (1) modifies nominal observations using generative image editing to vary different environmental factors, and (2) predicts performance under each variation using a policy-specific anomaly detector executed on edited observations. Experiments across 500+ hardware trials in twelve off-nominal conditions for visuomotor diffusion policies demonstrate that RoboART predicts performance degradation with high accuracy (less than 0.19 average difference between predicted and real success rates). We also demonstrate how predictive red teaming enables targeted data collection: fine-tuning with data collected under conditions predicted to be adverse boosts baseline performance by 2-7x.

Ola Shorinwa, Jiankai Sun, Mac Schwager et al.

We present SIREN for registration of multi-robot Gaussian Splatting (GSplat) maps, with zero access to camera poses, images, and inter-map transforms for initialization or fusion of local submaps. To realize these capabilities, SIREN harnesses the versatility and robustness of semantics in three critical ways to derive a rigorous registration pipeline for multi-robot GSplat maps. First, SIREN utilizes semantics to identify feature-rich regions of the local maps where the registration problem is better posed, eliminating the need for any initialization which is generally required in prior work. Second, SIREN identifies candidate correspondences between Gaussians in the local maps using robust semantic features, constituting the foundation for robust geometric optimization, coarsely aligning 3D Gaussian primitives extracted from the local maps. Third, this key step enables subsequent photometric refinement of the transformation between the submaps, where SIREN leverages novel-view synthesis in GSplat maps along with a semantics-based image filter to compute a high-accuracy non-rigid transformation for the generation of a high-fidelity fused map. We demonstrate the superior performance of SIREN compared to competing baselines across a range of real-world datasets, and in particular, across the most widely-used robot hardware platforms, including a manipulator, drone, and quadruped. In our experiments, SIREN achieves about 90x smaller rotation errors, 300x smaller translation errors, and 44x smaller scale errors in the most challenging scenes, where competing methods struggle. We will release the code and provide a link to the project page after the review process.

Pierre Sermanet, Anirudha Majumdar, Alex Irpan et al.

Until recently, robotics safety research was predominantly about collision avoidance and hazard reduction in the immediate vicinity of a robot. Since the advent of large vision and language models (VLMs), robots are now also capable of higher-level semantic scene understanding and natural language interactions with humans. Despite their known vulnerabilities (e.g. hallucinations or jail-breaking), VLMs are being handed control of robots capable of physical contact with the real world. This can lead to dangerous behaviors, making semantic safety for robots a matter of immediate concern. Our contributions in this paper are two fold: first, to address these emerging risks, we release the ASIMOV Benchmark, a large-scale and comprehensive collection of datasets for evaluating and improving semantic safety of foundation models serving as robot brains. Our data generation recipe is highly scalable: by leveraging text and image generation techniques, we generate undesirable situations from real-world visual scenes and human injury reports from hospitals. Secondly, we develop a framework to automatically generate robot constitutions from real-world data to steer a robot's behavior using Constitutional AI mechanisms. We propose a novel auto-amending process that is able to introduce nuances in written rules of behavior; this can lead to increased alignment with human preferences on behavior desirability and safety. We explore trade-offs between generality and specificity across a diverse set of constitutions of different lengths, and demonstrate that a robot is able to effectively reject unconstitutional actions. We measure a top alignment rate of 84.3% on the ASIMOV Benchmark using generated constitutions, outperforming no-constitution baselines and human-written constitutions. Data is available at asimov-benchmark.github.io

Lihan Zha, Apurva Badithela, Michael Zhang et al.

Generalist imitation learning policies trained on large datasets show great promise for solving diverse manipulation tasks. However, to ensure generalization to different conditions, policies need to be trained with data collected across a large set of environmental factor variations (e.g., camera pose, table height, distractors) $-$ a prohibitively expensive undertaking, if done exhaustively. We introduce a principled method for deciding what data to collect and how much to collect for each factor by constructing factored scaling curves (FSC), which quantify how policy performance varies as data scales along individual or paired factors. These curves enable targeted data acquisition for the most influential factor combinations within a given budget. We evaluate the proposed method through extensive simulated and real-world experiments, across both training-from-scratch and fine-tuning settings, and show that it boosts success rates in real-world tasks in new environments by up to 26% over existing data-collection strategies. We further demonstrate how factored scaling curves can effectively guide data collection using an offline metric, without requiring real-world evaluation at scale.

Tenny Yin, Zhiting Mei, Tao Sun et al.

Language-instructed active object localization is a critical challenge for robots, requiring efficient exploration of partially observable environments. However, state-of-the-art approaches either struggle to generalize beyond demonstration datasets (e.g., imitation learning methods) or fail to generate physically grounded actions (e.g., VLMs). To address these limitations, we introduce WoMAP (World Models for Active Perception): a recipe for training open-vocabulary object localization policies that: (i) uses a Gaussian Splatting-based real-to-sim-to-real pipeline for scalable data generation without the need for expert demonstrations, (ii) distills dense rewards signals from open-vocabulary object detectors, and (iii) leverages a latent world model for dynamics and rewards prediction to ground high-level action proposals at inference time. Rigorous simulation and hardware experiments demonstrate WoMAP's superior performance in a broad range of zero-shot object localization tasks, with more than 9x and 2x higher success rates compared to VLM and diffusion policy baselines, respectively. Further, we show that WoMAP achieves strong generalization and sim-to-real transfer on a TidyBot.

Gemini Robotics Team, Krzysztof Choromanski, Coline Devin et al.

Generative world models hold significant potential for simulating interactions with visuomotor policies in varied environments. Frontier video models can enable generation of realistic observations and environment interactions in a scalable and general manner. However, the use of video models in robotics has been limited primarily to in-distribution evaluations, i.e., scenarios that are similar to ones used to train the policy or fine-tune the base video model. In this report, we demonstrate that video models can be used for the entire spectrum of policy evaluation use cases in robotics: from assessing nominal performance to out-of-distribution (OOD) generalization, and probing physical and semantic safety. We introduce a generative evaluation system built upon a frontier video foundation model (Veo). The system is optimized to support robot action conditioning and multi-view consistency, while integrating generative image-editing and multi-view completion to synthesize realistic variations of real-world scenes along multiple axes of generalization. We demonstrate that the system preserves the base capabilities of the video model to enable accurate simulation of scenes that have been edited to include novel interaction objects, novel visual backgrounds, and novel distractor objects. This fidelity enables accurately predicting the relative performance of different policies in both nominal and OOD conditions, determining the relative impact of different axes of generalization on policy performance, and performing red teaming of policies to expose behaviors that violate physical or semantic safety constraints. We validate these capabilities through 1600+ real-world evaluations of eight Gemini Robotics policy checkpoints and five tasks for a bimanual manipulator.

Apurva Badithela, David Snyder, Lihan Zha et al.

Rapid progress in imitation learning, foundation models, and large-scale datasets has led to robot manipulation policies that generalize to a wide-range of tasks and environments. However, rigorous evaluation of these policies remains a challenge. Typically in practice, robot policies are often evaluated on a small number of hardware trials without any statistical assurances. We present SureSim, a framework to augment large-scale simulation with relatively small-scale real-world testing to provide reliable inferences on the real-world performance of a policy. Our key idea is to formalize the problem of combining real and simulation evaluations as a prediction-powered inference problem, in which a small number of paired real and simulation evaluations are used to rectify bias in large-scale simulation. We then leverage non-asymptotic mean estimation algorithms to provide confidence intervals on mean policy performance. Using physics-based simulation, we evaluate both diffusion policy and multi-task fine-tuned \(π_0\) on a joint distribution of objects and initial conditions, and find that our approach saves over \(20-25\%\) of hardware evaluation effort to achieve similar bounds on policy performance.

Zhiting Mei, Ola Shorinwa, Anirudha Majumdar

Semantic distillation in radiance fields has spurred significant advances in open-vocabulary robot policies, e.g., in manipulation and navigation, founded on pretrained semantics from large vision models. While prior work has demonstrated the effectiveness of visual-only semantic features (e.g., DINO and CLIP) in Gaussian Splatting and neural radiance fields, the potential benefit of geometry-grounding in distilled fields remains an open question. In principle, visual-geometry features seem very promising for spatial tasks such as pose estimation, prompting the question: Do geometry-grounded semantic features offer an edge in distilled fields? Specifically, we ask three critical questions: First, does spatial-grounding produce higher-fidelity geometry-aware semantic features? We find that image features from geometry-grounded backbones contain finer structural details compared to their counterparts. Secondly, does geometry-grounding improve semantic object localization? We observe no significant difference in this task. Thirdly, does geometry-grounding enable higher-accuracy radiance field inversion? Given the limitations of prior work and their lack of semantics integration, we propose a novel framework SPINE for inverting radiance fields without an initial guess, consisting of two core components: coarse inversion using distilled semantics, and fine inversion using photometric-based optimization. Surprisingly, we find that the pose estimation accuracy decreases with geometry-grounded features. Our results suggest that visual-only features offer greater versatility for a broader range of downstream tasks, although geometry-grounded features contain more geometric detail. Notably, our findings underscore the necessity of future research on effective strategies for geometry-grounding that augment the versatility and performance of pretrained semantic features.

Bowen Feng, Zhiting Mei, Baiang Li et al.

While autonomous driving (AD) stacks struggle with decision making under partial observability and real-world complexity, human drivers are capable of commonsense reasoning to make near-optimal decisions with limited information. Recent work has attempted to leverage finetuned Vision-Language Models (VLMs) for trajectory planning at inference time to emulate human behavior. Despite their success in benchmark evaluations, these methods are often impractical to deploy (a 70B parameter VLM inference at merely 8 tokens per second requires more than 160G of memory), and their monolithic network structure prohibits safety decomposition. To bridge this gap, we propose VLM-Embedded Reasoning for autonomous Driving (VERDI), a training-time framework that distills the reasoning process and commonsense knowledge of VLMs into the AD stack. VERDI augments modular differentiable end-to-end (e2e) AD models by aligning intermediate module outputs at the perception, prediction, and planning stages with text features explaining the driving reasoning process produced by VLMs. By encouraging alignment in latent space, VERDI enables the modular AD stack to internalize structured reasoning, without incurring the inference-time costs of large VLMs. We demonstrate the effectiveness of our method on the NuScenes dataset and find that VERDI outperforms existing e2e methods that do not embed reasoning by 10% in $\ell_{2}$ distance, while maintaining high inference speed.

Allen Z. Ren, Brian Ichter, Anirudha Majumdar

Large language models (LLMs) have exhibited remarkable reasoning and planning capabilities. Most prior work in this area has used LLMs to reason through steps from an initial to a goal state or criterion, thereby effectively reasoning in a forward direction. Nonetheless, many planning problems exhibit an inherent asymmetry such that planning backward from the goal is significantly easier -- for example, if there are bottlenecks close to the goal. We take inspiration from this observation and demonstrate that this bias holds for LLM planning as well: planning performance in one direction correlates with the planning complexity of the problem in that direction. However, our experiments also reveal systematic biases which lead to poor planning in the backward direction. With this knowledge, we propose a backward planning algorithm for LLMs that first flips the problem and then plans forward in the flipped problem. This helps avoid the backward bias, generate more diverse candidate plans, and exploit asymmetries between the forward and backward directions in planning problems -- we find that combining planning in both directions with self-verification improves the overall planning success rates by 4-24% in three planning domains.

Zhiting Mei, Tenny Yin, Micah Baker et al.

Recent advances in generative video models have led to significant breakthroughs in high-fidelity video synthesis, specifically in controllable video generation where the generated video is conditioned on text and action inputs, e.g., in instruction-guided video editing and world modeling in robotics. Despite these exceptional capabilities, controllable video models often hallucinate - generating future video frames that are misaligned with physical reality - which raises serious concerns in many tasks such as robot policy evaluation and planning. However, state-of-the-art video models lack the ability to assess and express their confidence, impeding hallucination mitigation. To rigorously address this challenge, we propose C3, an uncertainty quantification (UQ) method for training continuous-scale calibrated controllable video models for dense confidence estimation at the subpatch level, precisely localizing the uncertainty in each generated video frame. Our UQ method introduces three core innovations to empower video models to estimate their uncertainty. First, our method develops a novel framework that trains video models for correctness and calibration via strictly proper scoring rules. Second, we estimate the video model's uncertainty in latent space, avoiding training instability and prohibitive training costs associated with pixel-space approaches. Third, we map the dense latent-space uncertainty to interpretable pixel-level uncertainty in the RGB space for intuitive visualization, providing high-resolution uncertainty heatmaps that identify untrustworthy regions. Through extensive experiments on large-scale robot learning datasets (Bridge and DROID) and real-world evaluations, we demonstrate that our method not only provides calibrated uncertainty estimates within the training distribution, but also enables effective out-of-distribution detection.

Zhiting Mei, Ola Shorinwa, Anirudha Majumdar

Generative video models demonstrate impressive text-to-video capabilities, spurring widespread adoption in many real-world applications. However, like large language models (LLMs), video generation models tend to hallucinate, producing plausible videos even when they are factually wrong. Although uncertainty quantification (UQ) of LLMs has been extensively studied in prior work, no UQ method for video models exists, raising critical safety concerns. To our knowledge, this paper represents the first work towards quantifying the uncertainty of video models. We present a framework for uncertainty quantification of generative video models, consisting of: (i) a metric for evaluating the calibration of video models based on robust rank correlation estimation with no stringent modeling assumptions; (ii) a black-box UQ method for video models (termed S-QUBED), which leverages latent modeling to rigorously decompose predictive uncertainty into its aleatoric and epistemic components; and (iii) a UQ dataset to facilitate benchmarking calibration in video models. By conditioning the generation task in the latent space, we disentangle uncertainty arising due to vague task specifications from that arising from lack of knowledge. Through extensive experiments on benchmark video datasets, we demonstrate that S-QUBED computes calibrated total uncertainty estimates that are negatively correlated with the task accuracy and effectively computes the aleatoric and epistemic constituents.

Abhishek Jindal, Dmitry Kalashnikov, Oscar Chang et al.

When AI interacts with the physical world -- as a robot or an assistive agent -- new safety challenges emerge beyond those of purely ``digital AI". In such interactions, the potential for physical harm is direct and immediate. How well do state-of-the-art foundation models understand common-sense facts about physical safety, e.g. that a box may be too heavy to lift, or that a hot cup of coffee should not be handed to a child? In this paper, our contributions are three-fold: first, we develop a highly scalable approach to continuous physical safety benchmarking of Embodied AI systems, grounded in real-world injury narratives and operational safety constraints. To probe multi-modal safety understanding, we turn these narratives and constraints into photorealistic images and videos capturing transitions from safe to unsafe states, using advanced generative models. Secondly, we comprehensively analyze the ability of major foundation models to perceive risks, reason about safety, and trigger interventions; this yields multi-faceted insights into their deployment readiness for safety-critical agentic applications. Finally, we develop a post-training paradigm to teach models to explicitly reason about embodiment-specific safety constraints provided through system instructions. The resulting models generate thinking traces that make safety reasoning interpretable and transparent, achieving state of the art performance in constraint satisfaction evaluations. The benchmark will be released at https://asimov-benchmark.github.io/v2

Anirudha Majumdar

This paper proposes deception as a mechanism for out-of-distribution (OOD) generalization: by learning data representations that make training data appear independent and identically distributed (iid) to an observer, we can identify stable features that eliminate spurious correlations and generalize to unseen domains. We refer to this principle as deceptive risk minimization (DRM) and instantiate it with a practical differentiable objective that simultaneously learns features that eliminate distribution shifts from the perspective of a detector based on conformal martingales while minimizing a task-specific loss. In contrast to domain adaptation or prior invariant representation learning methods, DRM does not require access to test data or a partitioning of training data into a finite number of data-generating domains. We demonstrate the efficacy of DRM on numerical experiments with concept shift and a simulated imitation learning setting with covariate shift in environments that a robot is deployed in.

Pierre Sermanet, Anirudha Majumdar, Vikas Sindhwani

Given the recent rate of progress in artificial intelligence (AI) and robotics, a tantalizing question is emerging: would robots controlled by emerging AI systems be strongly aligned with human values? In this work, we propose a scalable way to probe this question by generating a benchmark spanning the key moments in 824 major pieces of science fiction literature (movies, tv, novels and scientific books) where an agent (AI or robot) made critical decisions (good or bad). We use a state-of-the-art LLM's recollection of each key moment to generate questions in similar situations, the decisions made by the agent, and alternative decisions it could have made (good or bad). We then measure an approximation of how well models align with human values on a set of human-voted answers. We also generate rules that can be automatically improved via an amendment process in order to generate the first Sci-Fi inspired constitutions for promoting ethical behavior in AIs and robots in the real world. Our first finding is that modern LLMs paired with constitutions turn out to be well-aligned with human values (95.8%), contrary to unsettling decisions typically made in Sci-Fi (only 21.2% alignment). Secondly, we find that generated constitutions substantially increase alignment compared to the base model (79.4% to 95.8%), and show resilience to an adversarial prompt setting (23.3% to 92.3%). Additionally, we find that those constitutions are among the top performers on the ASIMOV Benchmark which is derived from real-world images and hospital injury reports. Sci-Fi-inspired constitutions are thus highly aligned and applicable in real-world situations. We release SciFi-Benchmark: a large-scale dataset to advance robot ethics and safety research. It comprises 9,056 questions and 53,384 answers generated through a novel LLM-introspection process, in addition to a smaller human-labeled evaluation set.

Michelle Ho, Alec Farid, Anirudha Majumdar

We are motivated by the problem of comparing the complexity of one robotic task relative to another. To this end, we define a notion of reduction that formalizes the following intuition: Task 1 reduces to Task 2 if we can efficiently transform any policy that solves Task 2 into a policy that solves Task 1. We further define a quantitative measure of the relative complexity between any two tasks for a given robot. We prove useful properties of our notion of reduction (e.g., reflexivity, transitivity, and antisymmetry) and relative complexity measure (e.g., nonnegativity and monotonicity). In addition, we propose practical algorithms for estimating the relative complexity measure. We illustrate our framework for comparing robotic tasks using (i) examples where one can analytically establish reductions, and (ii) reinforcement learning examples where the proposed algorithm can estimate the relative complexity between tasks.

Alec Farid, David Snyder, Allen Z. Ren et al.

We are motivated by the problem of performing failure prediction for safety-critical robotic systems with high-dimensional sensor observations (e.g., vision). Given access to a black-box control policy (e.g., in the form of a neural network) and a dataset of training environments, we present an approach for synthesizing a failure predictor with guaranteed bounds on false-positive and false-negative errors. In order to achieve this, we utilize techniques from Probably Approximately Correct (PAC)-Bayes generalization theory. In addition, we present novel class-conditional bounds that allow us to trade-off the relative rates of false-positive vs. false-negative errors. We propose algorithms that train failure predictors (that take as input the history of sensor observations) by minimizing our theoretical error bounds. We demonstrate the resulting approach using extensive simulation and hardware experiments for vision-based navigation with a drone and grasping objects with a robotic manipulator equipped with a wrist-mounted RGB-D camera. These experiments illustrate the ability of our approach to (1) provide strong bounds on failure prediction error rates (that closely match empirical error rates), and (2) improve safety by predicting failures.

Anirudha Majumdar, Zhiting Mei, Vincent Pacelli

Our goal is to develop theory and algorithms for establishing fundamental limits on performance imposed by a robot's sensors for a given task. In order to achieve this, we define a quantity that captures the amount of task-relevant information provided by a sensor. Using a novel version of the generalized Fano inequality from information theory, we demonstrate that this quantity provides an upper bound on the highest achievable expected reward for one-step decision making tasks. We then extend this bound to multi-step problems via a dynamic programming approach. We present algorithms for numerically computing the resulting bounds, and demonstrate our approach on three examples: (i) the lava problem from the literature on partially observable Markov decision processes, (ii) an example with continuous state and observation spaces corresponding to a robot catching a freely-falling object, and (iii) obstacle avoidance using a depth sensor with non-Gaussian noise. We demonstrate the ability of our approach to establish strong limits on achievable performance for these problems by comparing our upper bounds with achievable lower bounds (computed by synthesizing or learning concrete control policies).

Kai-Chieh Hsu, Allen Z. Ren, Duy Phuong Nguyen et al.

Safety is a critical component of autonomous systems and remains a challenge for learning-based policies to be utilized in the real world. In particular, policies learned using reinforcement learning often fail to generalize to novel environments due to unsafe behavior. In this paper, we propose Sim-to-Lab-to-Real to bridge the reality gap with a probabilistically guaranteed safety-aware policy distribution. To improve safety, we apply a dual policy setup where a performance policy is trained using the cumulative task reward and a backup (safety) policy is trained by solving the Safety Bellman Equation based on Hamilton-Jacobi (HJ) reachability analysis. In Sim-to-Lab transfer, we apply a supervisory control scheme to shield unsafe actions during exploration; in Lab-to-Real transfer, we leverage the Probably Approximately Correct (PAC)-Bayes framework to provide lower bounds on the expected performance and safety of policies in unseen environments. Additionally, inheriting from the HJ reachability analysis, the bound accounts for the expectation over the worst-case safety in each environment. We empirically study the proposed framework for ego-vision navigation in two types of indoor environments with varying degrees of photorealism. We also demonstrate strong generalization performance through hardware experiments in real indoor spaces with a quadrupedal robot. See https://sites.google.com/princeton.edu/sim-to-lab-to-real for supplementary material.

Abhinav Agarwal, Sushant Veer, Allen Z. Ren et al.

We are motivated by the problem of learning policies for robotic systems with rich sensory inputs (e.g., vision) in a manner that allows us to guarantee generalization to environments unseen during training. We provide a framework for providing such generalization guarantees by leveraging a finite dataset of real-world environments in combination with a (potentially inaccurate) generative model of environments. The key idea behind our approach is to utilize the generative model in order to implicitly specify a prior over policies. This prior is updated using the real-world dataset of environments by minimizing an upper bound on the expected cost across novel environments derived via Probably Approximately Correct (PAC)-Bayes generalization theory. We demonstrate our approach on two simulated systems with nonlinear/hybrid dynamics and rich sensing modalities: (i) quadrotor navigation with an onboard vision sensor, and (ii) grasping objects using a depth sensor. Comparisons with prior work demonstrate the ability of our approach to obtain stronger generalization guarantees by utilizing generative models. We also present hardware experiments for validating our bounds for the grasping task.

Ali Ekin Gurgen, Anirudha Majumdar, Sushant Veer

This paper presents an approach for learning motion planners that are accompanied with probabilistic guarantees of success on new environments that hold uniformly for any disturbance to the robot's dynamics within an admissible set. We achieve this by bringing together tools from generalization theory and robust control. First, we curate a library of motion primitives where the robustness of each primitive is characterized by an over-approximation of the forward reachable set, i.e., a "funnel". Then, we optimize probably approximately correct (PAC)-Bayes generalization bounds for training our planner to compose these primitives such that the entire funnels respect the problem specification. We demonstrate the ability of our approach to provide strong guarantees on two simulated examples: (i) navigation of an autonomous vehicle under external disturbances on a five-lane highway with multiple vehicles, and (ii) navigation of a drone across an obstacle field in the presence of wind disturbances.

Vincent Pacelli, Anirudha Majumdar

The rapid development of affordable and compact high-fidelity sensors (e.g., cameras and LIDAR) allows robots to construct detailed estimates of their states and environments. However, the availability of such rich sensor information introduces two technical challenges: (i) the lack of analytic sensing models, which makes it difficult to design controllers that are robust to sensor failures, and (ii) the computational expense of processing the high-dimensional sensor information in real time. This paper addresses these challenges using the theory of differential privacy, which allows us to (i) design controllers with bounded sensitivity to errors in state estimates, and (ii) bound the amount of state information used for control (i.e., to impose bounded rationality). The resulting framework approximates the separation principle and allows us to derive an upper-bound on the cost incurred with a faulty state estimator in terms of three quantities: the cost incurred using a perfect state estimator, the magnitude of state estimation errors, and the level of differential privacy. We demonstrate the efficacy of our framework numerically on different robotics problems, including nonlinear system stabilization and motion planning.

Allen Z. Ren, Anirudha Majumdar

Our goal is to train control policies that generalize well to unseen environments. Inspired by the Distributionally Robust Optimization (DRO) framework, we propose DRAGEN - Distributionally Robust policy learning via Adversarial Generation of ENvironments - for iteratively improving robustness of policies to realistic distribution shifts by generating adversarial environments. The key idea is to learn a generative model for environments whose latent variables capture cost-predictive and realistic variations in environments. We perform DRO with respect to a Wasserstein ball around the empirical distribution of environments by generating realistic adversarial environments via gradient ascent on the latent space. We demonstrate strong Out-of-Distribution (OoD) generalization in simulation for (i) swinging up a pendulum with onboard vision and (ii) grasping realistic 3D objects. Grasping experiments on hardware demonstrate better sim2real performance compared to domain randomization.

Alec Farid, Sushant Veer, Divyanshu Pachisia et al.

Our goal is to perform out-of-distribution (OOD) detection, i.e., to detect when a robot is operating in environments drawn from a different distribution than the ones used to train the robot. We leverage Probably Approximately Correct (PAC)-Bayes theory to train a policy with a guaranteed bound on performance on the training distribution. Our idea for OOD detection relies on the following intuition: violation of the performance bound on test environments provides evidence that the robot is operating OOD. We formalize this via statistical techniques based on p-values and concentration inequalities. The approach provides guaranteed confidence bounds on OOD detection including bounds on both the false positive and false negative rates of the detector and is task-driven and only sensitive to changes that impact the robot's performance. We demonstrate our approach in simulation and hardware for a grasping task using objects with unfamiliar shapes or poses and a drone performing vision-based obstacle avoidance in environments with wind disturbances and varied obstacle densities. Our examples demonstrate that we can perform task-driven OOD detection within just a handful of trials.

Naman Agarwal, Elad Hazan, Anirudha Majumdar et al.

We consider the setting of iterative learning control, or model-based policy learning in the presence of uncertain, time-varying dynamics. In this setting, we propose a new performance metric, planning regret, which replaces the standard stochastic uncertainty assumptions with worst case regret. Based on recent advances in non-stochastic control, we design a new iterative algorithm for minimizing planning regret that is more robust to model mismatch and uncertainty. We provide theoretical and empirical evidence that the proposed algorithm outperforms existing methods on several benchmarks.

Alec Farid, Anirudha Majumdar

We are motivated by the problem of providing strong generalization guarantees in the context of meta-learning. Existing generalization bounds are either challenging to evaluate or provide vacuous guarantees in even relatively simple settings. We derive a probably approximately correct (PAC) bound for gradient-based meta-learning using two different generalization frameworks in order to deal with the qualitatively different challenges of generalization at the "base" and "meta" levels. We employ bounds for uniformly stable algorithms at the base level and bounds from the PAC-Bayes framework at the meta level. The result of this approach is a novel PAC bound that is tighter when the base learner adapts quickly, which is precisely the goal of meta-learning. We show that our bound provides a tighter guarantee than other bounds on a toy non-convex problem on the unit sphere and a text-based classification example. We also present a practical regularization scheme motivated by the bound in settings where the bound is loose and demonstrate improved performance over baseline techniques.

Udaya Ghai, David Snyder, Anirudha Majumdar et al.

We consider the problem of generating maximally adversarial disturbances for a given controller assuming only blackbox access to it. We propose an online learning approach to this problem that \emph{adaptively} generates disturbances based on control inputs chosen by the controller. The goal of the disturbance generator is to minimize \emph{regret} versus a benchmark disturbance-generating policy class, i.e., to maximize the cost incurred by the controller as well as possible compared to the best possible disturbance generator \emph{in hindsight} (chosen from a benchmark policy class). In the setting where the dynamics are linear and the costs are quadratic, we formulate our problem as an online trust region (OTR) problem with memory and present a new online learning algorithm (\emph{MOTR}) for this problem. We prove that this method competes with the best disturbance generator in hindsight (chosen from a rich class of benchmark policies that includes linear-dynamical disturbance generating policies). We demonstrate our approach on two simulated examples: (i) synthetically generated linear systems, and (ii) generating wind disturbances for the popular PX4 controller in the AirSim simulator. On these examples, we demonstrate that our approach outperforms several baseline approaches, including $H_{\infty}$ disturbance generation and gradient-based methods.

Christine Allen-Blanchette, Sushant Veer, Anirudha Majumdar et al.

The dominant paradigms for video prediction rely on opaque transition models where neither the equations of motion nor the underlying physical quantities of the system are easily inferred. The equations of motion, as defined by Newton's second law, describe the time evolution of a physical system state and can therefore be applied toward the determination of future system states. In this paper, we introduce a video prediction model where the equations of motion are explicitly constructed from learned representations of the underlying physical quantities. To achieve this, we simultaneously learn a low-dimensional state representation and system Lagrangian. The kinetic and potential energy terms of the Lagrangian are distinctly modelled and the low-dimensional equations of motion are explicitly constructed using the Euler-Lagrange equations. We demonstrate the efficacy of this approach for video prediction on image sequences rendered in modified OpenAI gym Pendulum-v0 and Acrobot environments.

Meghan Booker, Anirudha Majumdar

Robots equipped with rich sensing modalities (e.g., RGB-D cameras) performing long-horizon tasks motivate the need for policies that are highly memory-efficient. State-of-the-art approaches for controlling robots often use memory representations that are excessively rich for the task or rely on hand-crafted tricks for memory efficiency. Instead, this work provides a general approach for jointly synthesizing memory representations and policies; the resulting policies actively seek to reduce memory requirements. Specifically, we present a reinforcement learning framework that leverages an implementation of the group LASSO regularization to synthesize policies that employ low-dimensional and task-centric memory representations. We demonstrate the efficacy of our approach with simulated examples including navigation in discrete and continuous spaces as well as vision-based indoor navigation set in a photo-realistic simulator. The results on these examples indicate that our method is capable of finding policies that rely only on low-dimensional memory representations, improving generalization, and actively reducing memory requirements.

Allen Z. Ren, Sushant Veer, Anirudha Majumdar

Control policies from imitation learning can often fail to generalize to novel environments due to imperfect demonstrations or the inability of imitation learning algorithms to accurately infer the expert's policies. In this paper, we present rigorous generalization guarantees for imitation learning by leveraging the Probably Approximately Correct (PAC)-Bayes framework to provide upper bounds on the expected cost of policies in novel environments. We propose a two-stage training method where a latent policy distribution is first embedded with multi-modal expert behavior using a conditional variational autoencoder, and then "fine-tuned" in new training environments to explicitly optimize the generalization bound. We demonstrate strong generalization bounds and their tightness relative to empirical performance in simulation for (i) grasping diverse mugs, (ii) planar pushing with visual feedback, and (iii) vision-based indoor navigation, as well as through hardware experiments for the two manipulation tasks.

Sushant Veer, Anirudha Majumdar

We present a novel algorithm -- convex natural evolutionary strategies (CoNES) -- for optimizing high-dimensional blackbox functions by leveraging tools from convex optimization and information geometry. CoNES is formulated as an efficiently-solvable convex program that adapts the evolutionary strategies (ES) gradient estimate to promote rapid convergence. The resulting algorithm is invariant to the parameterization of the belief distribution. Our numerical results demonstrate that CoNES vastly outperforms conventional blackbox optimization methods on a suite of functions used for benchmarking blackbox optimizers. Furthermore, CoNES demonstrates the ability to converge faster than conventional blackbox methods on a selection of OpenAI's MuJoCo reinforcement learning tasks for locomotion.

Anoopkumar Sonar, Vincent Pacelli, Anirudha Majumdar

A fundamental challenge in reinforcement learning is to learn policies that generalize beyond the operating domains experienced during training. In this paper, we approach this challenge through the following invariance principle: an agent must find a representation such that there exists an action-predictor built on top of this representation that is simultaneously optimal across all training domains. Intuitively, the resulting invariant policy enhances generalization by finding causes of successful actions. We propose a novel learning algorithm, Invariant Policy Optimization (IPO), that implements this principle and learns an invariant policy during training. We compare our approach with standard policy gradient methods and demonstrate significant improvements in generalization performance on unseen domains for linear quadratic regulator and grid-world problems, and an example where a robot must learn to open doors with varying physical properties.