Franziska Meier

Arjun Majumdar, Karmesh Yadav, Sergio Arnaud et al. · meta-ai

We present the largest and most comprehensive empirical study of pre-trained visual representations (PVRs) or visual 'foundation models' for Embodied AI. First, we curate CortexBench, consisting of 17 different tasks spanning locomotion, navigation, dexterous, and mobile manipulation. Next, we systematically evaluate existing PVRs and find that none are universally dominant. To study the effect of pre-training data size and diversity, we combine over 4,000 hours of egocentric videos from 7 different sources (over 4.3M images) and ImageNet to train different-sized vision transformers using Masked Auto-Encoding (MAE) on slices of this data. Contrary to inferences from prior work, we find that scaling dataset size and diversity does not improve performance universally (but does so on average). Our largest model, named VC-1, outperforms all prior PVRs on average but does not universally dominate either. Next, we show that task- or domain-specific adaptation of VC-1 leads to substantial gains, with VC-1 (adapted) achieving competitive or superior performance than the best known results on all of the benchmarks in CortexBench. Finally, we present real-world hardware experiments, in which VC-1 and VC-1 (adapted) outperform the strongest pre-existing PVR. Overall, this paper presents no new techniques but a rigorous systematic evaluation, a broad set of findings about PVRs (that in some cases, refute those made in narrow domains in prior work), and open-sourced code and models (that required over 10,000 GPU-hours to train) for the benefit of the research community.

Karl Pertsch, Ruta Desai, Vikash Kumar et al.

We propose an approach for semantic imitation, which uses demonstrations from a source domain, e.g. human videos, to accelerate reinforcement learning (RL) in a different target domain, e.g. a robotic manipulator in a simulated kitchen. Instead of imitating low-level actions like joint velocities, our approach imitates the sequence of demonstrated semantic skills like "opening the microwave" or "turning on the stove". This allows us to transfer demonstrations across environments (e.g. real-world to simulated kitchen) and agent embodiments (e.g. bimanual human demonstration to robotic arm). We evaluate on three challenging cross-domain learning problems and match the performance of demonstration-accelerated RL approaches that require in-domain demonstrations. In a simulated kitchen environment, our approach learns long-horizon robot manipulation tasks, using less than 3 minutes of human video demonstrations from a real-world kitchen. This enables scaling robot learning via the reuse of demonstrations, e.g. collected as human videos, for learning in any number of target domains.

Sneha Silwal, Karmesh Yadav, Tingfan Wu et al.

We present a large empirical investigation on the use of pre-trained visual representations (PVRs) for training downstream policies that execute real-world tasks. Our study involves five different PVRs, each trained for five distinct manipulation or indoor navigation tasks. We performed this evaluation using three different robots and two different policy learning paradigms. From this effort, we can arrive at three insights: 1) the performance trends of PVRs in the simulation are generally indicative of their trends in the real world, 2) the use of PVRs enables a first-of-its-kind result with indoor ImageNav (zero-shot transfer to a held-out scene in the real world), and 3) the benefits from variations in PVRs, primarily data-augmentation and fine-tuning, also transfer to the real-world performance. See project website for additional details and visuals.

Michael Chang, Alyssa L. Dayan, Franziska Meier et al.

Object rearrangement is a challenge for embodied agents because solving these tasks requires generalizing across a combinatorially large set of configurations of entities and their locations. Worse, the representations of these entities are unknown and must be inferred from sensory percepts. We present a hierarchical abstraction approach to uncover these underlying entities and achieve combinatorial generalization from unstructured visual inputs. By constructing a factorized transition graph over clusters of entity representations inferred from pixels, we show how to learn a correspondence between intervening on states of entities in the agent's model and acting on objects in the environment. We use this correspondence to develop a method for control that generalizes to different numbers and configurations of objects, which outperforms current offline deep RL methods when evaluated on simulated rearrangement tasks.

Andrew Szot, Amy Zhang, Dhruv Batra et al.

How well do reward functions learned with inverse reinforcement learning (IRL) generalize? We illustrate that state-of-the-art IRL algorithms, which maximize a maximum-entropy objective, learn rewards that overfit to the demonstrations. Such rewards struggle to provide meaningful rewards for states not covered by the demonstrations, a major detriment when using the reward to learn policies in new situations. We introduce BC-IRL a new inverse reinforcement learning method that learns reward functions that generalize better when compared to maximum-entropy IRL approaches. In contrast to the MaxEnt framework, which learns to maximize rewards around demonstrations, BC-IRL updates reward parameters such that the policy trained with the new reward matches the expert demonstrations better. We show that BC-IRL learns rewards that generalize better on an illustrative simple task and two continuous robotic control tasks, achieving over twice the success rate of baselines in challenging generalization settings.

Sarah Bechtle, Ludovic Righetti, Franziska Meier

Being able to seamlessly generalize across different tasks is fundamental for robots to act in our world. However, learning representations that generalize quickly to new scenarios is still an open research problem in reinforcement learning. In this paper we present a framework to meta-learn the critic for gradient-based policy learning. Concretely, we propose a model-based bi-level optimization algorithm that updates the critics parameters such that the policy that is learned with the updated critic gets closer to solving the meta-training tasks. We illustrate that our algorithm leads to learned critics that resemble the ground truth Q function for a given task. Finally, after meta-training, the learned critic can be used to learn new policies for new unseen task and environment settings via model-free policy gradient optimization, without requiring a model. We present results that show the generalization capabilities of our learned critic to new tasks and dynamics when used to learn a new policy in a new scenario.

Matthias De Lange, Hamid Eghbalzadeh, Reuben Tan et al.

In egocentric action recognition a single population model is typically trained and subsequently embodied on a head-mounted device, such as an augmented reality headset. While this model remains static for new users and environments, we introduce an adaptive paradigm of two phases, where after pretraining a population model, the model adapts on-device and online to the user's experience. This setting is highly challenging due to the change from population to user domain and the distribution shifts in the user's data stream. Coping with the latter in-stream distribution shifts is the focus of continual learning, where progress has been rooted in controlled benchmarks but challenges faced in real-world applications often remain unaddressed. We introduce EgoAdapt, a benchmark for real-world egocentric action recognition that facilitates our two-phased adaptive paradigm, and real-world challenges naturally occur in the egocentric video streams from Ego4d, such as long-tailed action distributions and large-scale classification over 2740 actions. We introduce an evaluation framework that directly exploits the user's data stream with new metrics to measure the adaptation gain over the population model, online generalization, and hindsight performance. In contrast to single-stream evaluation in existing works, our framework proposes a meta-evaluation that aggregates the results from 50 independent user streams. We provide an extensive empirical study for finetuning and experience replay.

Jianing Yang, Alexander Sax, Kevin J. Liang et al.

Multi-view 3D reconstruction remains a core challenge in computer vision, particularly in applications requiring accurate and scalable representations across diverse perspectives. Current leading methods such as DUSt3R employ a fundamentally pairwise approach, processing images in pairs and necessitating costly global alignment procedures to reconstruct from multiple views. In this work, we propose Fast 3D Reconstruction (Fast3R), a novel multi-view generalization to DUSt3R that achieves efficient and scalable 3D reconstruction by processing many views in parallel. Fast3R's Transformer-based architecture forwards N images in a single forward pass, bypassing the need for iterative alignment. Through extensive experiments on camera pose estimation and 3D reconstruction, Fast3R demonstrates state-of-the-art performance, with significant improvements in inference speed and reduced error accumulation. These results establish Fast3R as a robust alternative for multi-view applications, offering enhanced scalability without compromising reconstruction accuracy.

Franziska Meier, Austin Wang, Giovanni Sutanto et al.

Building differentiable simulations of physical processes has recently received an increasing amount of attention. Specifically, some efforts develop differentiable robotic physics engines motivated by the computational benefits of merging rigid body simulations with modern differentiable machine learning libraries. Here, we present a library that focuses on the ability to combine data driven methods with analytical rigid body computations. More concretely, our library \emph{Differentiable Robot Models} implements both \emph{differentiable} and \emph{learnable} models of the kinematics and dynamics of robots in Pytorch. The source-code is available at \url{https://github.com/facebookresearch/differentiable-robot-model}

Sayna Ebrahimi, Franziska Meier, Roberto Calandra et al.

Continual learning aims to learn new tasks without forgetting previously learned ones. We hypothesize that representations learned to solve each task in a sequence have a shared structure while containing some task-specific properties. We show that shared features are significantly less prone to forgetting and propose a novel hybrid continual learning framework that learns a disjoint representation for task-invariant and task-specific features required to solve a sequence of tasks. Our model combines architecture growth to prevent forgetting of task-specific skills and an experience replay approach to preserve shared skills. We demonstrate our hybrid approach is effective in avoiding forgetting and show it is superior to both architecture-based and memory-based approaches on class incrementally learning of a single dataset as well as a sequence of multiple datasets in image classification. Our code is available at \url{https://github.com/facebookresearch/Adversarial-Continual-Learning}.

Sarah Bechtle, Artem Molchanov, Yevgen Chebotar et al.

Typically, loss functions, regularization mechanisms and other important aspects of training parametric models are chosen heuristically from a limited set of options. In this paper, we take the first step towards automating this process, with the view of producing models which train faster and more robustly. Concretely, we present a meta-learning method for learning parametric loss functions that can generalize across different tasks and model architectures. We develop a pipeline for meta-training such loss functions, targeted at maximizing the performance of the model trained under them. The loss landscape produced by our learned losses significantly improves upon the original task-specific losses in both supervised and reinforcement learning tasks. Furthermore, we show that our meta-learning framework is flexible enough to incorporate additional information at meta-train time. This information shapes the learned loss function such that the environment does not need to provide this information during meta-test time. We make our code available at https://sites.google.com/view/mlthree.

Mido Assran, Adrien Bardes, David Fan et al. · meta-ai

A major challenge for modern AI is to learn to understand the world and learn to act largely by observation. This paper explores a self-supervised approach that combines internet-scale video data with a small amount of interaction data (robot trajectories), to develop models capable of understanding, predicting, and planning in the physical world. We first pre-train an action-free joint-embedding-predictive architecture, V-JEPA 2, on a video and image dataset comprising over 1 million hours of internet video. V-JEPA 2 achieves strong performance on motion understanding (77.3 top-1 accuracy on Something-Something v2) and state-of-the-art performance on human action anticipation (39.7 recall-at-5 on Epic-Kitchens-100) surpassing previous task-specific models. Additionally, after aligning V-JEPA 2 with a large language model, we demonstrate state-of-the-art performance on multiple video question-answering tasks at the 8 billion parameter scale (e.g., 84.0 on PerceptionTest, 76.9 on TempCompass). Finally, we show how self-supervised learning can be applied to robotic planning tasks by post-training a latent action-conditioned world model, V-JEPA 2-AC, using less than 62 hours of unlabeled robot videos from the Droid dataset. We deploy V-JEPA 2-AC zero-shot on Franka arms in two different labs and enable picking and placing of objects using planning with image goals. Notably, this is achieved without collecting any data from the robots in these environments, and without any task-specific training or reward. This work demonstrates how self-supervised learning from web-scale data and a small amount of robot interaction data can yield a world model capable of planning in the physical world.

Pascale Fung, Yoram Bachrach, Asli Celikyilmaz et al.

This paper describes our research on AI agents embodied in visual, virtual or physical forms, enabling them to interact with both users and their environments. These agents, which include virtual avatars, wearable devices, and robots, are designed to perceive, learn and act within their surroundings, which makes them more similar to how humans learn and interact with the environments as compared to disembodied agents. We propose that the development of world models is central to reasoning and planning of embodied AI agents, allowing these agents to understand and predict their environment, to understand user intentions and social contexts, thereby enhancing their ability to perform complex tasks autonomously. World modeling encompasses the integration of multimodal perception, planning through reasoning for action and control, and memory to create a comprehensive understanding of the physical world. Beyond the physical world, we also propose to learn the mental world model of users to enable better human-agent collaboration.

Sergio Arnaud, Paul McVay, Ada Martin et al. · mit

We present LOCATE 3D, a model for localizing objects in 3D scenes from referring expressions like "the small coffee table between the sofa and the lamp." LOCATE 3D sets a new state-of-the-art on standard referential grounding benchmarks and showcases robust generalization capabilities. Notably, LOCATE 3D operates directly on sensor observation streams (posed RGB-D frames), enabling real-world deployment on robots and AR devices. Key to our approach is 3D-JEPA, a novel self-supervised learning (SSL) algorithm applicable to sensor point clouds. It takes as input a 3D pointcloud featurized using 2D foundation models (CLIP, DINO). Subsequently, masked prediction in latent space is employed as a pretext task to aid the self-supervised learning of contextualized pointcloud features. Once trained, the 3D-JEPA encoder is finetuned alongside a language-conditioned decoder to jointly predict 3D masks and bounding boxes. Additionally, we introduce LOCATE 3D DATASET, a new dataset for 3D referential grounding, spanning multiple capture setups with over 130K annotations. This enables a systematic study of generalization capabilities as well as a stronger model.

Ayush Jain, Alexander Swerdlow, Yuzhou Wang et al.

Progress in 3D vision-language learning has been hindered by the scarcity of large-scale 3D datasets. We introduce UniVLG, a unified architecture for 2D and 3D vision-language understanding that bridges the gap between existing 2D-centric models and the rich 3D sensory data available in embodied systems. Our approach initializes most model weights from pre-trained 2D models and trains on both 2D and 3D vision-language data. We propose a novel language-conditioned mask decoder shared across 2D and 3D modalities to ground objects effectively in both RGB and RGB-D images, outperforming box-based approaches. To further reduce the domain gap between 2D and 3D, we incorporate 2D-to-3D lifting strategies, enabling UniVLG to utilize 2D data to enhance 3D performance. With these innovations, our model achieves state-of-the-art performance across multiple 3D vision-language grounding tasks, demonstrating the potential of transferring advances from 2D vision-language learning to the data-constrained 3D domain. Furthermore, co-training on both 2D and 3D data enhances performance across modalities without sacrificing 2D capabilities. By removing the reliance on 3D mesh reconstruction and ground-truth object proposals, UniVLG sets a new standard for realistic, embodied-aligned evaluation. Code and additional visualizations are available at https://univlg.github.io .

Ang Cao, Sergio Arnaud, Oleksandr Maksymets et al.

3D vision-language grounding faces a fundamental data bottleneck: while 2D models train on billions of images, 3D models have access to only thousands of labeled scenes--a six-order-of-magnitude gap that severely limits performance. We introduce $\textbf{LIFT-GS}$, a practical distillation technique that overcomes this limitation by using differentiable rendering to bridge 3D and 2D supervision. LIFT-GS predicts 3D Gaussian representations from point clouds and uses them to render predicted language-conditioned 3D masks into 2D views, enabling supervision from 2D foundation models (SAM, CLIP, LLaMA) without requiring any 3D annotations. This render-supervised formulation enables end-to-end training of complete encoder-decoder architectures and is inherently model-agnostic. LIFT-GS achieves state-of-the-art results with $25.7\%$ mAP on open-vocabulary instance segmentation (vs. $20.2\%$ prior SOTA) and consistent $10-30\%$ improvements on referential grounding tasks. Remarkably, pretraining effectively multiplies fine-tuning datasets by 2X, demonstrating strong scaling properties that suggest 3D VLG currently operates in a severely data-scarce regime. Project page: https://liftgs.github.io

Shagun Sodhani, Franziska Meier, Joelle Pineau et al.

In reinforcement learning (RL), when defining a Markov Decision Process (MDP), the environment dynamics is implicitly assumed to be stationary. This assumption of stationarity, while simplifying, can be unrealistic in many scenarios. In the continual reinforcement learning scenario, the sequence of tasks is another source of nonstationarity. In this work, we propose to examine this continual reinforcement learning setting through the block contextual MDP (BC-MDP) framework, which enables us to relax the assumption of stationarity. This framework challenges RL algorithms to handle both nonstationarity and rich observation settings and, by additionally leveraging smoothness properties, enables us to study generalization bounds for this setting. Finally, we take inspiration from adaptive control to propose a novel algorithm that addresses the challenges introduced by this more realistic BC-MDP setting, allows for zero-shot adaptation at evaluation time, and achieves strong performance on several nonstationary environments.

Todor Davchev, Sarah Bechtle, Subramanian Ramamoorthy et al.

Inverse reinforcement learning is a paradigm motivated by the goal of learning general reward functions from demonstrated behaviours. Yet the notion of generality for learnt costs is often evaluated in terms of robustness to various spatial perturbations only, assuming deployment at fixed speeds of execution. However, this is impractical in the context of robotics and building, time-invariant solutions is of crucial importance. In this work, we propose a formulation that allows us to 1) vary the length of execution by learning time-invariant costs, and 2) relax the temporal alignment requirements for learning from demonstration. We apply our method to two different types of cost formulations and evaluate their performance in the context of learning reward functions for simulated placement and peg in hole tasks executed on a 7DoF Kuka IIWA arm. Our results show that our approach enables learning temporally invariant rewards from misaligned demonstration that can also generalise spatially to out of distribution tasks.

Andrew Szot, Alex Clegg, Eric Undersander et al.

We introduce Habitat 2.0 (H2.0), a simulation platform for training virtual robots in interactive 3D environments and complex physics-enabled scenarios. We make comprehensive contributions to all levels of the embodied AI stack - data, simulation, and benchmark tasks. Specifically, we present: (i) ReplicaCAD: an artist-authored, annotated, reconfigurable 3D dataset of apartments (matching real spaces) with articulated objects (e.g. cabinets and drawers that can open/close); (ii) H2.0: a high-performance physics-enabled 3D simulator with speeds exceeding 25,000 simulation steps per second (850x real-time) on an 8-GPU node, representing 100x speed-ups over prior work; and, (iii) Home Assistant Benchmark (HAB): a suite of common tasks for assistive robots (tidy the house, prepare groceries, set the table) that test a range of mobile manipulation capabilities. These large-scale engineering contributions allow us to systematically compare deep reinforcement learning (RL) at scale and classical sense-plan-act (SPA) pipelines in long-horizon structured tasks, with an emphasis on generalization to new objects, receptacles, and layouts. We find that (1) flat RL policies struggle on HAB compared to hierarchical ones; (2) a hierarchy with independent skills suffers from 'hand-off problems', and (3) SPA pipelines are more brittle than RL policies.

Kalesha Bullard, Douwe Kiela, Franziska Meier et al.

Effective communication is an important skill for enabling information exchange in multi-agent settings and emergent communication is now a vibrant field of research, with common settings involving discrete cheap-talk channels. Since, by definition, these settings involve arbitrary encoding of information, typically they do not allow for the learned protocols to generalize beyond training partners. In contrast, in this work, we present a novel problem setting and the Quasi-Equivalence Discovery (QED) algorithm that allows for zero-shot coordination (ZSC), i.e., discovering protocols that can generalize to independently trained agents. Real world problem settings often contain costly communication channels, e.g., robots have to physically move their limbs, and a non-uniform distribution over intents. We show that these two factors lead to unique optimal ZSC policies in referential games, where agents use the energy cost of the messages to communicate intent. Other-Play was recently introduced for learning optimal ZSC policies, but requires prior access to the symmetries of the problem. Instead, QED can iteratively discovers the symmetries in this setting and converges to the optimal ZSC policy.

Joanne Truong, Denis Yarats, Tianyu Li et al.

Recent work has shown results on learning navigation policies for idealized cylinder agents in simulation and transferring them to real wheeled robots. Deploying such navigation policies on legged robots can be challenging due to their complex dynamics, and the large dynamical difference between cylinder agents and legged systems. In this work, we learn hierarchical navigation policies that account for the low-level dynamics of legged robots, such as maximum speed, slipping, contacts, and learn to successfully navigate cluttered indoor environments. To enable transfer of policies learned in simulation to new legged robots and hardware, we learn dynamics-aware navigation policies across multiple robots with robot-specific embeddings. The learned embedding is optimized on new robots, while the rest of the policy is kept fixed, allowing for quick adaptation. We train our policies across three legged robots in simulation - 2 quadrupeds (A1, AlienGo) and a hexapod (Daisy). At test time, we study the performance of our learned policy on two new legged robots in simulation (Laikago, 4-legged Daisy), and one real-world quadrupedal robot (A1). Our experiments show that our learned policy can sample-efficiently generalize to previously unseen robots, and enable sim-to-real transfer of navigation policies for legged robots.

Sarah Bechtle, Neha Das, Franziska Meier

Humans have impressive generalization capabilities when it comes to manipulating objects and tools in completely novel environments. These capabilities are, at least partially, a result of humans having internal models of their bodies and any grasped object. How to learn such body schemas for robots remains an open problem. In this work, we develop an self-supervised approach that can extend a robot's kinematic model when grasping an object from visual latent representations. Our framework comprises two components: (1) we present a multi-modal keypoint detector: an autoencoder architecture trained by fusing proprioception and vision to predict visual key points on an object; (2) we show how we can use our learned keypoint detector to learn an extension of the kinematic chain by regressing virtual joints from the predicted visual keypoints. Our evaluation shows that our approach learns to consistently predict visual keypoints on objects in the manipulator's hand, and thus can easily facilitate learning an extended kinematic chain to include the object grasped in various configurations, from a few seconds of visual data. Finally we show that this extended kinematic chain lends itself for object manipulation tasks such as placing a grasped object and present experiments in simulation and on hardware.

Sarah Bechtle, Bilal Hammoud, Akshara Rai et al.

Learning for model based control can be sample-efficient and generalize well, however successfully learning models and controllers that represent the problem at hand can be challenging for complex tasks. Using inaccurate models for learning can lead to sub-optimal solutions, that are unlikely to perform well in practice. In this work, we present a learning approach which iterates between model learning and data collection and leverages forward model prediction error for learning control. We show how using the controller's prediction as input to a forward model can create a differentiable connection between the controller and the model, allowing us to formulate a loss in the state space. This lets us include forward model prediction error during controller learning and we show that this creates a loss objective that significantly improves learning on different motor control tasks. We provide empirical and theoretical results that show the benefits of our method and present evaluations in simulation for learning control on a 7 DoF manipulator and an underactuated 12 DoF quadruped. We show that our approach successfully learns controllers for challenging motor control tasks involving contact switching.

Kalesha Bullard, Franziska Meier, Douwe Kiela et al.

Effective communication is an important skill for enabling information exchange and cooperation in multi-agent settings. Indeed, emergent communication is now a vibrant field of research, with common settings involving discrete cheap-talk channels. One limitation of this setting is that it does not allow for the emergent protocols to generalize beyond the training partners. Furthermore, so far emergent communication has primarily focused on the use of symbolic channels. In this work, we extend this line of work to a new modality, by studying agents that learn to communicate via actuating their joints in a 3D environment. We show that under realistic assumptions, a non-uniform distribution of intents and a common-knowledge energy cost, these agents can find protocols that generalize to novel partners. We also explore and analyze specific difficulties associated with finding these solutions in practice. Finally, we propose and evaluate initial training improvements to address these challenges, involving both specific training curricula and providing the latent feature that can be coordinated on during training.

Neha Das, Sarah Bechtle, Todor Davchev et al.

Scaling model-based inverse reinforcement learning (IRL) to real robotic manipulation tasks with unknown dynamics remains an open problem. The key challenges lie in learning good dynamics models, developing algorithms that scale to high-dimensional state-spaces and being able to learn from both visual and proprioceptive demonstrations. In this work, we present a gradient-based inverse reinforcement learning framework that utilizes a pre-trained visual dynamics model to learn cost functions when given only visual human demonstrations. The learned cost functions are then used to reproduce the demonstrated behavior via visual model predictive control. We evaluate our framework on hardware on two basic object manipulation tasks.

Tianyu Li, Roberto Calandra, Deepak Pathak et al.

Hierarchical learning has been successful at learning generalizable locomotion skills on walking robots in a sample-efficient manner. However, the low-dimensional "latent" action used to communicate between two layers of the hierarchy is typically user-designed. In this work, we present a fully-learned hierarchical framework, that is capable of jointly learning the low-level controller and the high-level latent action space. Once this latent space is learned, we plan over continuous latent actions in a model-predictive control fashion, using a learned high-level dynamics model. This framework generalizes to multiple robots, and we present results on a Daisy hexapod simulation, A1 quadruped simulation, and Daisy robot hardware. We compare a range of learned hierarchical approaches from literature, and show that our framework outperforms baselines on multiple tasks and two simulations. In addition to learning approaches, we also compare to inverse-kinematics (IK) acting on desired robot motion, and show that our fully-learned framework outperforms IK in adverse settings on both A1 and Daisy simulations. On hardware, we show the Daisy hexapod achieve multiple locomotion tasks, in an unstructured outdoor setting, with only 2000 hardware samples, reinforcing the robustness and sample-efficiency of our approach.

Todor Davchev, Kevin Sebastian Luck, Michael Burke et al.

Manipulation skills involving contact and friction are inherent to many robotics tasks. Using the class of motor primitives for peg-in-hole like insertions, we study how robots can learn such skills. Dynamic Movement Primitives (DMP) are a popular way of extracting such policies through behaviour cloning (BC) but can struggle in the context of insertion. Policy adaptation strategies such as residual learning can help improve the overall performance of policies in the context of contact-rich manipulation. However, it is not clear how to best do this with DMPs. As a result, we consider several possible ways for adapting a DMP formulation and propose ``residual Learning from Demonstration`` (rLfD), a framework that combines DMPs with Reinforcement Learning (RL) to learn a residual correction policy. Our evaluations suggest that applying residual learning directly in task space and operating on the full pose of the robot can significantly improve the overall performance of DMPs. We show that rLfD offers a gentle to the joints solution that improves the task success and generalisation of DMPs \rb{and enables transfer to different geometries and frictions through few-shot task adaptation}. The proposed framework is evaluated on a set of tasks. A simulated robot and a physical robot have to successfully insert pegs, gears and plugs into their respective sockets. Other material and videos accompanying this paper are provided at https://sites.google.com/view/rlfd/.

Giovanni Sutanto, Katharina Rombach, Yevgen Chebotar et al.

Robots need to be able to adapt to unexpected changes in the environment such that they can autonomously succeed in their tasks. However, hand-designing feedback models for adaptation is tedious, if at all possible, making data-driven methods a promising alternative. In this paper we introduce a full framework for learning feedback models for reactive motion planning. Our pipeline starts by segmenting demonstrations of a complete task into motion primitives via a semi-automated segmentation algorithm. Then, given additional demonstrations of successful adaptation behaviors, we learn initial feedback models through learning from demonstrations. In the final phase, a sample-efficient reinforcement learning algorithm fine-tunes these feedback models for novel task settings through few real system interactions. We evaluate our approach on a real anthropomorphic robot in learning a tactile feedback task.

Kristen Morse, Neha Das, Yixin Lin et al.

Being able to quickly adapt to changes in dynamics is paramount in model-based control for object manipulation tasks. In order to influence fast adaptation of the inverse dynamics model's parameters, data efficiency is crucial. Given observed data, a key element to how an optimizer updates model parameters is the loss function. In this work, we propose to apply meta-learning to learn structured, state-dependent loss functions during a meta-training phase. We then replace standard losses with our learned losses during online adaptation tasks. We evaluate our proposed approach on inverse dynamics learning tasks, both in simulation and on real hardware data. In both settings, the structured and state-dependent learned losses improve online adaptation speed, when compared to standard, state-independent loss functions.

Giovanni Sutanto, Austin S. Wang, Yixin Lin et al.

The recursive Newton-Euler Algorithm (RNEA) is a popular technique for computing the dynamics of robots. RNEA can be framed as a differentiable computational graph, enabling the dynamics parameters of the robot to be learned from data via modern auto-differentiation toolboxes. However, the dynamics parameters learned in this manner can be physically implausible. In this work, we incorporate physical constraints in the learning by adding structure to the learned parameters. This results in a framework that can learn physically plausible dynamics via gradient descent, improving the training speed as well as generalization of the learned dynamics models. We evaluate our method on real-time inverse dynamics control tasks on a 7 degree of freedom robot arm, both in simulation and on the real robot. Our experiments study a spectrum of structure added to the parameters of the differentiable RNEA algorithm, and compare their performance and generalization.

Edward Grefenstette, Brandon Amos, Denis Yarats et al.

Many (but not all) approaches self-qualifying as "meta-learning" in deep learning and reinforcement learning fit a common pattern of approximating the solution to a nested optimization problem. In this paper, we give a formalization of this shared pattern, which we call GIMLI, prove its general requirements, and derive a general-purpose algorithm for implementing similar approaches. Based on this analysis and algorithm, we describe a library of our design, higher, which we share with the community to assist and enable future research into these kinds of meta-learning approaches. We end the paper by showcasing the practical applications of this framework and library through illustrative experiments and ablation studies which they facilitate.

Tianyu Li, Nathan Lambert, Roberto Calandra et al.

Learning to locomote to arbitrary goals on hardware remains a challenging problem for reinforcement learning. In this paper, we present a hierarchical learning framework that improves sample-efficiency and generalizability of locomotion skills on real-world robots. Our approach divides the problem of goal-oriented locomotion into two sub-problems: learning diverse primitives skills, and using model-based planning to sequence these skills. We parametrize our primitives as cyclic movements, improving sample-efficiency of learning on a 18 degrees of freedom robot. Then, we learn coarse dynamics models over primitive cycles and use them in a model predictive control framework. This allows us to learn to walk to arbitrary goals up to 12m away, after about two hours of training from scratch on hardware. Our results on a Daisy hexapod hardware and simulation demonstrate the efficacy of our approach at reaching distant targets, in different environments and with sensory noise.

Sarah Bechtle, Yixin Lin, Akshara Rai et al.

Curiosity as a means to explore during reinforcement learning problems has recently become very popular. However, very little progress has been made in utilizing curiosity for learning control. In this work, we propose a model-based reinforcement learning (MBRL) framework that combines Bayesian modeling of the system dynamics with curious iLQR, an iterative LQR approach that considers model uncertainty. During trajectory optimization the curious iLQR attempts to minimize both the task-dependent cost and the uncertainty in the dynamics model. We demonstrate the approach on reaching tasks with 7-DoF manipulators in simulation and on a real robot. Our experiments show that MBRL with curious iLQR reaches desired end-effector targets more reliably and with less system rollouts when learning a new task from scratch, and that the learned model generalizes better to new reaching tasks.

Behnoosh Parsa, Keshav Rajasekaran, Franziska Meier et al.

One of the challenges in model-based control of stochastic dynamical systems is that the state transition dynamics are involved, and it is not easy or efficient to make good-quality predictions of the states. Moreover, there are not many representational models for the majority of autonomous systems, as it is not easy to build a compact model that captures the entire dynamical subtleties and uncertainties. In this work, we present a hierarchical Bayesian linear regression model with local features to learn the dynamics of a micro-robotic system as well as two simpler examples, consisting of a stochastic mass-spring damper and a stochastic double inverted pendulum on a cart. The model is hierarchical since we assume non-stationary priors for the model parameters. These non-stationary priors make the model more flexible by imposing priors on the priors of the model. To solve the maximum likelihood (ML) problem for this hierarchical model, we use the variational expectation maximization (EM) algorithm, and enhance the procedure by introducing hidden target variables. The algorithm yields parsimonious model structures, and consistently provides fast and accurate predictions for all our examples involving large training and test sets. This demonstrates the effectiveness of the method in learning stochastic dynamics, which makes it suitable for future use in a paradigm, such as model-based reinforcement learning, to compute optimal control policies in real time.

Akshara Rai, Rika Antonova, Franziska Meier et al.

Learning for control can acquire controllers for novel robotic tasks, paving the path for autonomous agents. Such controllers can be expert-designed policies, which typically require tuning of parameters for each task scenario. In this context, Bayesian optimization (BO) has emerged as a promising approach for automatically tuning controllers. However, when performing BO on hardware for high-dimensional policies, sample-efficiency can be an issue. Here, we develop an approach that utilizes simulation to map the original parameter space into a domain-informed space. During BO, similarity between controllers is now calculated in this transformed space. Experiments on the ATRIAS robot hardware and another bipedal robot simulation show that our approach succeeds at sample-efficiently learning controllers for multiple robots. Another question arises: What if the simulation significantly differs from hardware? To answer this, we create increasingly approximate simulators and study the effect of increasing simulation-hardware mismatch on the performance of Bayesian optimization. We also compare our approach to other approaches from literature, and find it to be more reliable, especially in cases of high mismatch. Our experiments show that our approach succeeds across different controller types, bipedal robot models and simulator fidelity levels, making it applicable to a wide range of bipedal locomotion problems.

Giovanni Sutanto, Zhe Su, Stefan Schaal et al.

In order to robustly execute a task under environmental uncertainty, a robot needs to be able to reactively adapt to changes arising in its environment. The environment changes are usually reflected in deviation from expected sensory traces. These deviations in sensory traces can be used to drive the motion adaptation, and for this purpose, a feedback model is required. The feedback model maps the deviations in sensory traces to the motion plan adaptation. In this paper, we develop a general data-driven framework for learning a feedback model from demonstrations. We utilize a variant of a radial basis function network structure --with movement phases as kernel centers-- which can generally be applied to represent any feedback models for movement primitives. To demonstrate the effectiveness of our framework, we test it on the task of scraping on a tilt board. In this task, we are learning a reactive policy in the form of orientation adaptation, based on deviations of tactile sensor traces. As a proof of concept of our method, we provide evaluations on an anthropomorphic robot. A video demonstrating our approach and its results can be seen in https://youtu.be/7Dx5imy1Kcw

Daniel Kappler, Franziska Meier, Nathan Ratliff et al.

Modern robotics is gravitating toward increasingly collaborative human robot interaction. Tools such as acceleration policies can naturally support the realization of reactive, adaptive, and compliant robots. These tools require us to model the system dynamics accurately -- a difficult task. The fundamental problem remains that simulation and reality diverge--we do not know how to accurately change a robot's state. Thus, recent research on improving inverse dynamics models has been focused on making use of machine learning techniques. Traditional learning techniques train on the actual realized accelerations, instead of the policy's desired accelerations, which is an indirect data source. Here we show how an additional training signal -- measured at the desired accelerations -- can be derived from a feedback control signal. This effectively creates a second data source for learning inverse dynamics models. Furthermore, we show how both the traditional and this new data source, can be used to train task-specific models of the inverse dynamics, when used independently or combined. We analyze the use of both data sources in simulation and demonstrate its effectiveness on a real-world robotic platform. We show that our system incrementally improves the learned inverse dynamics model, and when using both data sources combined converges more consistently and faster.

Arunkumar Byravan, Felix Leeb, Franziska Meier et al.

In this work, we present an approach to deep visuomotor control using structured deep dynamics models. Our deep dynamics model, a variant of SE3-Nets, learns a low-dimensional pose embedding for visuomotor control via an encoder-decoder structure. Unlike prior work, our dynamics model is structured: given an input scene, our network explicitly learns to segment salient parts and predict their pose-embedding along with their motion modeled as a change in the pose space due to the applied actions. We train our model using a pair of point clouds separated by an action and show that given supervision only in the form of point-wise data associations between the frames our network is able to learn a meaningful segmentation of the scene along with consistent poses. We further show that our model can be used for closed-loop control directly in the learned low-dimensional pose space, where the actions are computed by minimizing error in the pose space using gradient-based methods, similar to traditional model-based control. We present results on controlling a Baxter robot from raw depth data in simulation and in the real world and compare against two baseline deep networks. Our method runs in real-time, achieves good prediction of scene dynamics and outperforms the baseline methods on multiple control runs. Video results can be found at: https://rse-lab.cs.washington.edu/se3-structured-deep-ctrl/

Franziska Meier, Daniel Kappler, Stefan Schaal

The promise of learning to learn for robotics rests on the hope that by extracting some information about the learning process itself we can speed up subsequent similar learning tasks. Here, we introduce a computationally efficient online meta-learning algorithm that builds and optimizes a memory model of the optimal learning rate landscape from previously observed gradient behaviors. While performing task specific optimization, this memory of learning rates predicts how to scale currently observed gradients. After applying the gradient scaling our meta-learner updates its internal memory based on the observed effect its prediction had. Our meta-learner can be combined with any gradient-based optimizer, learns on the fly and can be transferred to new optimization tasks. In our evaluations we show that our meta-learning algorithm speeds up learning of MNIST classification and a variety of learning control tasks, either in batch or online learning settings.

Daniel Kappler, Franziska Meier, Jan Issac et al.

We address the challenging problem of robotic grasping and manipulation in the presence of uncertainty. This uncertainty is due to noisy sensing, inaccurate models and hard-to-predict environment dynamics. We quantify the importance of continuous, real-time perception and its tight integration with reactive motion generation methods in dynamic manipulation scenarios. We compare three different systems that are instantiations of the most common architectures in the field: (i) a traditional sense-plan-act approach that is still widely used, (ii) a myopic controller that only reacts to local environment dynamics and (iii) a reactive planner that integrates feedback control and motion optimization. All architectures rely on the same components for real-time perception and reactive motion generation to allow a quantitative evaluation. We extensively evaluate the systems on a real robotic platform in four scenarios that exhibit either a challenging workspace geometry or a dynamic environment. In 333 experiments, we quantify the robustness and accuracy that is due to integrating real-time feedback at different time scales in a reactive motion generation system. We also report on the lessons learned for system building.

Franziska Meier, Stefan Schaal

Dynamic Movement Primitives have successfully been used to realize imitation learning, trial-and-error learning, reinforce- ment learning, movement recognition and segmentation and control. Because of this they have become a popular represen- tation for motor primitives. In this work, we showcase how DMPs can be reformulated as a probabilistic linear dynamical system with control inputs. Through this probabilistic repre- sentation of DMPs, algorithms such as Kalman filtering and smoothing are directly applicable to perform inference on pro- prioceptive sensor measurements during execution. We show that inference in this probabilistic model automatically leads to a feedback term to online modulate the execution of a DMP. Furthermore, we show how inference allows us to measure the likelihood that we are successfully executing a given motion primitive. In this context, we show initial results of using the probabilistic model to detect execution failures on a simulated movement primitive dataset.

Akshara Rai, Giovanni Sutanto, Stefan Schaal et al.

With the advancement of robotics, machine learning, and machine perception, increasingly more robots will enter human environments to assist with daily tasks. However, dynamically-changing human environments requires reactive motion plans. Reactivity can be accomplished through replanning, e.g. model-predictive control, or through a reactive feedback policy that modifies on-going behavior in response to sensory events. In this paper, we investigate how to use machine learning to add reactivity to a previously learned nominal skilled behavior. We approach this by learning a reactive modification term for movement plans represented by nonlinear differential equations. In particular, we use dynamic movement primitives (DMPs) to represent a skill and a neural network to learn a reactive policy from human demonstrations. We use the well explored domain of obstacle avoidance for robot manipulation as a test bed. Our approach demonstrates how a neural network can be combined with physical insights to ensure robust behavior across different obstacle settings and movement durations. Evaluations on an anthropomorphic robotic system demonstrate the effectiveness of our work.

Nathan Ratliff, Franziska Meier, Daniel Kappler et al.

It has long been hoped that model-based control will improve tracking performance while maintaining or increasing compliance. This hope hinges on having or being able to estimate an accurate inverse dynamics model. As a result, substantial effort has gone into modeling and estimating dynamics (error) models. Most recent research has focused on learning the true inverse dynamics using data points mapping observed accelerations to the torques used to generate them. Unfortunately, if the initial tracking error is bad, such learning processes may train substantially off-distribution to predict well on actual observed acceleration rather then the desired accelerations. This work takes a different approach. We define a class of gradient-based online learning algorithms we term Direct Online Optimization for Modeling Errors in Dynamics (DOOMED) that directly minimize an objective measuring the divergence between actual and desired accelerations. Our objective is defined in terms of the true system's unknown dynamics and is therefore impossible to evaluate. However, we show that its gradient is measurable online from system data. We develop a novel adaptive control approach based on running online learning to directly correct (inverse) dynamics errors in real time using the data stream from the robot to accurately achieve desired accelerations during execution.

Manuel Wüthrich, Cristina Garcia Cifuentes, Sebastian Trimpe et al.

Many sensors, such as range, sonar, radar, GPS and visual devices, produce measurements which are contaminated by outliers. This problem can be addressed by using fat-tailed sensor models, which account for the possibility of outliers. Unfortunately, all estimation algorithms belonging to the family of Gaussian filters (such as the widely-used extended Kalman filter and unscented Kalman filter) are inherently incompatible with such fat-tailed sensor models. The contribution of this paper is to show that any Gaussian filter can be made compatible with fat-tailed sensor models by applying one simple change: Instead of filtering with the physical measurement, we propose to filter with a pseudo measurement obtained by applying a feature function to the physical measurement. We derive such a feature function which is optimal under some conditions. Simulation results show that the proposed method can effectively handle measurement outliers and allows for robust filtering in both linear and nonlinear systems.

Franziska Meier, Philipp Hennig, Stefan Schaal

Locally weighted regression was created as a nonparametric learning method that is computationally efficient, can learn from very large amounts of data and add data incrementally. An interesting feature of locally weighted regression is that it can work with spatially varying length scales, a beneficial property, for instance, in control problems. However, it does not provide a generative model for function values and requires training and test data to be generated identically, independently. Gaussian (process) regression, on the other hand, provides a fully generative model without significant formal requirements on the distribution of training data, but has much higher computational cost and usually works with one global scale per input dimension. Using a localising function basis and approximate inference techniques, we take Gaussian (process) regression to increasingly localised properties and toward the same computational complexity class as locally weighted regression.