Aviv Tamar

Shie Mannor, Aviv Tamar

Reinforcement learning (RL) has demonstrated great potential, but is currently full of overhyping and pipe dreams. We point to some difficulties with current research which we feel are endemic to the direction taken by the community. To us, the current direction is not likely to lead to "deployable" RL: RL that works in practice and can work in practical situations yet still is economically viable. We also propose a potential fix to some of the difficulties of the field.

Ev Zisselman, Itai Lavie, Daniel Soudry et al.

We study zero-shot generalization in reinforcement learning-optimizing a policy on a set of training tasks to perform well on a similar but unseen test task. To mitigate overfitting, previous work explored different notions of invariance to the task. However, on problems such as the ProcGen Maze, an adequate solution that is invariant to the task visualization does not exist, and therefore invariance-based approaches fail. Our insight is that learning a policy that effectively $\textit{explores}$ the domain is harder to memorize than a policy that maximizes reward for a specific task, and therefore we expect such learned behavior to generalize well; we indeed demonstrate this empirically on several domains that are difficult for invariance-based approaches. Our $\textit{Explore to Generalize}$ algorithm (ExpGen) builds on this insight: we train an additional ensemble of agents that optimize reward. At test time, either the ensemble agrees on an action, and we generalize well, or we take exploratory actions, which generalize well and drive us to a novel part of the state space, where the ensemble may potentially agree again. We show that our approach is the state-of-the-art on tasks of the ProcGen challenge that have thus far eluded effective generalization, yielding a success rate of $83\%$ on the Maze task and $74\%$ on Heist with $200$ training levels. ExpGen can also be combined with an invariance based approach to gain the best of both worlds, setting new state-of-the-art results on ProcGen.

Idan Shenfeld, Zhang-Wei Hong, Aviv Tamar et al.

Learning from rewards (i.e., reinforcement learning or RL) and learning to imitate a teacher (i.e., teacher-student learning) are two established approaches for solving sequential decision-making problems. To combine the benefits of these different forms of learning, it is common to train a policy to maximize a combination of reinforcement and teacher-student learning objectives. However, without a principled method to balance these objectives, prior work used heuristics and problem-specific hyperparameter searches to balance the two objectives. We present a $\textit{principled}$ approach, along with an approximate implementation for $\textit{dynamically}$ and $\textit{automatically}$ balancing when to follow the teacher and when to use rewards. The main idea is to adjust the importance of teacher supervision by comparing the agent's performance to the counterfactual scenario of the agent learning without teacher supervision and only from rewards. If using teacher supervision improves performance, the importance of teacher supervision is increased and otherwise it is decreased. Our method, $\textit{Teacher Guided Reinforcement Learning}$ (TGRL), outperforms strong baselines across diverse domains without hyper-parameter tuning.

Tal Daniel, Aviv Tamar

We propose a new representation of visual data that disentangles object position from appearance. Our method, termed Deep Latent Particles (DLP), decomposes the visual input into low-dimensional latent ``particles'', where each particle is described by its spatial location and features of its surrounding region. To drive learning of such representations, we follow a VAE-based approach and introduce a prior for particle positions based on a spatial-softmax architecture, and a modification of the evidence lower bound loss inspired by the Chamfer distance between particles. We demonstrate that our DLP representations are useful for downstream tasks such as unsupervised keypoint (KP) detection, image manipulation, and video prediction for scenes composed of multiple dynamic objects. In addition, we show that our probabilistic interpretation of the problem naturally provides uncertainty estimates for particle locations, which can be used for model selection, among other tasks. Videos and code are available: https://taldatech.github.io/deep-latent-particles-web/

Tal Daniel, Aviv Tamar

We propose a new object-centric video prediction algorithm based on the deep latent particle (DLP) representation. In comparison to existing slot- or patch-based representations, DLPs model the scene using a set of keypoints with learned parameters for properties such as position and size, and are both efficient and interpretable. Our method, deep dynamic latent particles (DDLP), yields state-of-the-art object-centric video prediction results on several challenging datasets. The interpretable nature of DDLP allows us to perform ``what-if'' generation -- predict the consequence of changing properties of objects in the initial frames, and DLP's compact structure enables efficient diffusion-based unconditional video generation. Videos, code and pre-trained models are available: https://taldatech.github.io/ddlp-web

Era Choshen, Aviv Tamar

In meta reinforcement learning (meta RL), an agent seeks a Bayes-optimal policy -- the optimal policy when facing an unknown task that is sampled from some known task distribution. Previous approaches tackled this problem by inferring a belief over task parameters, using variational inference methods. Motivated by recent successes of contrastive learning approaches in RL, such as contrastive predictive coding (CPC), we investigate whether contrastive methods can be used for learning Bayes-optimal behavior. We begin by proving that representations learned by CPC are indeed sufficient for Bayes optimality. Based on this observation, we propose a simple meta RL algorithm that uses CPC in lieu of variational belief inference. Our method, ContraBAR, achieves comparable performance to state-of-the-art in domains with state-based observation and circumvents the computational toll of future observation reconstruction, enabling learning in domains with image-based observations. It can also be combined with image augmentations for domain randomization and used seamlessly in both online and offline meta RL settings.

Zohar Rimon, Aviv Tamar, Gilad Adler

In meta reinforcement learning (meta RL), an agent learns from a set of training tasks how to quickly solve a new task, drawn from the same task distribution. The optimal meta RL policy, a.k.a. the Bayes-optimal behavior, is well defined, and guarantees optimal reward in expectation, taken with respect to the task distribution. The question we explore in this work is how many training tasks are required to guarantee approximately optimal behavior with high probability. Recent work provided the first such PAC analysis for a model-free setting, where a history-dependent policy was learned from the training tasks. In this work, we propose a different approach: directly learn the task distribution, using density estimation techniques, and then train a policy on the learned task distribution. We show that our approach leads to bounds that depend on the dimension of the task distribution. In particular, in settings where the task distribution lies in a low-dimensional manifold, we extend our analysis to use dimensionality reduction techniques and account for such structure, obtaining significantly better bounds than previous work, which strictly depend on the number of states and actions. The key of our approach is the regularization implied by the kernel density estimation method. We further demonstrate that this regularization is useful in practice, when `plugged in' the state-of-the-art VariBAD meta RL algorithm.

Yarin Perry, Felipe Vieira Frujeri, Chaim Hoch et al.

Routing is, arguably, the most fundamental task in computer networking, and the most extensively studied one. A key challenge for routing in real-world environments is the need to contend with uncertainty about future traffic demands. We present a new approach to routing under demand uncertainty: tackling this challenge as stochastic optimization, and employing deep learning to learn complex patterns in traffic demands. We show that our method provably converges to the global optimum in well-studied theoretical models of multicommodity flow. We exemplify the practical usefulness of our approach by zooming in on the real-world challenge of traffic engineering (TE) on wide-area networks (WANs). Our extensive empirical evaluation on real-world traffic and network topologies establishes that our approach's TE quality almost matches that of an (infeasible) omniscient oracle, outperforming previously proposed approaches, and also substantially lowers runtimes.

Orr Krupnik, Elisei Shafer, Tom Jurgenson et al.

Adaptable models could greatly benefit robotic agents operating in the real world, allowing them to deal with novel and varying conditions. While approaches such as Bayesian inference are well-studied frameworks for adapting models to evidence, we build on recent advances in deep generative models which have greatly affected many areas of robotics. Harnessing modern GPU acceleration, we investigate how to quickly adapt the sample generation of neural network models to observations in robotic tasks. We propose a simple and general method that is applicable to various deep generative models and robotic environments. The key idea is to quickly fine-tune the model by fitting it to generated samples matching the observed evidence, using the cross-entropy method. We show that our method can be applied to both autoregressive models and variational autoencoders, and demonstrate its usability in object shape inference from grasping, inverse kinematics calculation, and point cloud completion.

Khashayar Rohanimanesh, Jake Metzger, William Richards et al.

Deep learning-based grasp prediction models have become an industry standard for robotic bin-picking systems. To maximize pick success, production environments are often equipped with several end-effector tools that can be swapped on-the-fly, based on the target object. Tool-change, however, takes time. Choosing the order of grasps to perform, and corresponding tool-change actions, can improve system throughput; this is the topic of our work. The main challenge in planning tool change is uncertainty - we typically cannot see objects in the bin that are currently occluded. Inspired by queuing and admission control problems, we model the problem as a Markov Decision Process (MDP), where the goal is to maximize expected throughput, and we pursue an approximate solution based on model predictive control, where at each time step we plan based only on the currently visible objects. Special to our method is the idea of void zones, which are geometrical boundaries in which an unknown object will be present, and therefore cannot be accounted for during planning. Our planning problem can be solved using integer linear programming (ILP). However, we find that an approximate solution based on sparse tree search yields near optimal performance at a fraction of the time. Another question that we explore is how to measure the performance of tool-change planning: we find that throughput alone can fail to capture delicate and smooth behavior, and propose a principled alternative. Finally, we demonstrate our algorithms on both synthetic and real world bin picking tasks.

Gal Leibovich, Guy Jacob, Or Avner et al.

We propose $\textit{iterative inversion}$ -- an algorithm for learning an inverse function without input-output pairs, but only with samples from the desired output distribution and access to the forward function. The key challenge is a $\textit{distribution shift}$ between the desired outputs and the outputs of an initial random guess, and we prove that iterative inversion can steer the learning correctly, under rather strict conditions on the function. We apply iterative inversion to learn control. Our input is a set of demonstrations of desired behavior, given as video embeddings of trajectories (without actions), and our method iteratively learns to imitate trajectories generated by the current policy, perturbed by random exploration noise. Our approach does not require rewards, and only employs supervised learning, which can be easily scaled to use state-of-the-art trajectory embedding techniques and policy representations. Indeed, with a VQ-VAE embedding, and a transformer-based policy, we demonstrate non-trivial continuous control on several tasks. Further, we report an improved performance on imitating diverse behaviors compared to reward based methods.

Nir Greshler, David Ben Eli, Carmel Rabinovitz et al.

The combination of Monte Carlo tree search and neural networks has revolutionized online planning. As neural network approximations are often imperfect, we ask whether uncertainty estimates about the network outputs could be used to improve planning. We develop a Bayesian planning approach that facilitates such uncertainty quantification, inspired by classical ideas from the meta-reasoning literature. We propose a Thompson sampling based algorithm for searching the tree of possible actions, for which we prove the first (to our knowledge) finite time Bayesian regret bound, and propose an efficient implementation for a restricted family of posterior distributions. In addition we propose a variant of the Bayes-UCB method applied to trees. Empirically, we demonstrate that on the ProcGen Maze and Leaper environments, when the uncertainty estimates are accurate but the neural network output is inaccurate, our Bayesian approach searches the tree much more effectively. In addition, we investigate whether popular uncertainty estimation methods are accurate enough to yield significant gains in planning. Our code is available at: https://github.com/nirgreshler/bayesian-online-planning.

Jianlan Luo, Eugen Solowjow, Chengtao Wen et al.

Precise robotic manipulation skills are desirable in many industrial settings, reinforcement learning (RL) methods hold the promise of acquiring these skills autonomously. In this paper, we explicitly consider incorporating operational space force/torque information into reinforcement learning; this is motivated by humans heuristically mapping perceived forces to control actions, which results in completing high-precision tasks in a fairly easy manner. Our approach combines RL with force/torque information by incorporating a proper operational space force controller; where we also exploit different ablations on processing this information. Moreover, we propose a neural network architecture that generalizes to reasonable variations of the environment. We evaluate our method on the open-source Siemens Robot Learning Challenge, which requires precise and delicate force-controlled behavior to assemble a tight-fit gear wheel set.

Elad Sarafian, Aviv Tamar, Sarit Kraus

We propose a policy improvement algorithm for Reinforcement Learning (RL) which is called Rerouted Behavior Improvement (RBI). RBI is designed to take into account the evaluation errors of the Q-function. Such errors are common in RL when learning the $Q$-value from finite past experience data. Greedy policies or even constrained policy optimization algorithms which ignore these errors may suffer from an improvement penalty (i.e. a negative policy improvement). To minimize the improvement penalty, the RBI idea is to attenuate rapid policy changes of low probability actions which were less frequently sampled. This approach is shown to avoid catastrophic performance degradation and reduce regret when learning from a batch of past experience. Through a two-armed bandit with Gaussian distributed rewards example, we show that it also increases data efficiency when the optimal action has a high variance. We evaluate RBI in two tasks in the Atari Learning Environment: (1) learning from observations of multiple behavior policies and (2) iterative RL. Our results demonstrate the advantage of RBI over greedy policies and other constrained policy optimization algorithms as a safe learning approach and as a general data efficient learning algorithm. An anonymous Github repository of our RBI implementation is found at https://github.com/eladsar/rbi.

Dan Haramati, Tal Daniel, Aviv Tamar

Manipulating objects is a hallmark of human intelligence, and an important task in domains such as robotics. In principle, Reinforcement Learning (RL) offers a general approach to learn object manipulation. In practice, however, domains with more than a few objects are difficult for RL agents due to the curse of dimensionality, especially when learning from raw image observations. In this work we propose a structured approach for visual RL that is suitable for representing multiple objects and their interaction, and use it to learn goal-conditioned manipulation of several objects. Key to our method is the ability to handle goals with dependencies between the objects (e.g., moving objects in a certain order). We further relate our architecture to the generalization capability of the trained agent, based on a theoretical result for compositional generalization, and demonstrate agents that learn with 3 objects but generalize to similar tasks with over 10 objects. Videos and code are available on the project website: https://sites.google.com/view/entity-centric-rl

Shelly Francis-Meretzki, Mirco Mutti, Yaniv Romano et al.

Recent advances in vision-language-action (VLA) models for robotics have highlighted the importance of reliable uncertainty quantification in sequential tasks. However, assessing and improving calibration in such settings remains mostly unexplored, especially when only partial trajectories are observed. In this work, we formulate sequential calibration for episodic tasks, where task-success confidence is produced along an episode, while success is determined at the end of it. We introduce a sequential extension of the Brier score and show that, for binary outcomes, its risk minimizer coincides with the VLA policy's value function. This connection bridges uncertainty calibration and reinforcement learning, enabling the use of temporal-difference (TD) value estimation as a principled calibration mechanism over time. We empirically show that TD calibration improves performance relative to the state-of-the-art on simulated and real-robot data. Interestingly, we show that when calibrated using TD, the VLA's single-step action probabilities can yield competitive uncertainty estimates, in contrast to recent findings that employed different calibration techniques.

Carl Qi, Dan Haramati, Tal Daniel et al.

Object manipulation is a common component of everyday tasks, but learning to manipulate objects from high-dimensional observations presents significant challenges. These challenges are heightened in multi-object environments due to the combinatorial complexity of the state space as well as of the desired behaviors. While recent approaches have utilized large-scale offline data to train models from pixel observations, achieving performance gains through scaling, these methods struggle with compositional generalization in unseen object configurations with constrained network and dataset sizes. To address these issues, we propose a novel behavioral cloning (BC) approach that leverages object-centric representations and an entity-centric Transformer with diffusion-based optimization, enabling efficient learning from offline image data. Our method first decomposes observations into an object-centric representation, which is then processed by our entity-centric Transformer that computes attention at the object level, simultaneously predicting object dynamics and the agent's actions. Combined with the ability of diffusion models to capture multi-modal behavior distributions, this results in substantial performance improvements in multi-object tasks and, more importantly, enables compositional generalization. We present BC agents capable of zero-shot generalization to tasks with novel compositions of objects and goals, including larger numbers of objects than seen during training. We provide video rollouts on our webpage: https://sites.google.com/view/ec-diffuser.

Tal Daniel, Carl Qi, Dan Haramati et al.

We introduce Latent Particle World Model (LPWM), a self-supervised object-centric world model scaled to real-world multi-object datasets and applicable in decision-making. LPWM autonomously discovers keypoints, bounding boxes, and object masks directly from video data, enabling it to learn rich scene decompositions without supervision. Our architecture is trained end-to-end purely from videos and supports flexible conditioning on actions, language, and image goals. LPWM models stochastic particle dynamics via a novel latent action module and achieves state-of-the-art results on diverse real-world and synthetic datasets. Beyond stochastic video modeling, LPWM is readily applicable to decision-making, including goal-conditioned imitation learning, as we demonstrate in the paper. Code, data, pre-trained models and video rollouts are available: https://taldatech.github.io/lpwm-web

Dan Haramati, Carl Qi, Tal Daniel et al.

We propose a hierarchical entity-centric framework for offline Goal-Conditioned Reinforcement Learning (GCRL) that combines subgoal decomposition with factored structure to solve long-horizon tasks in domains with multiple entities. Achieving long-horizon goals in complex environments remains a core challenge in Reinforcement Learning (RL). Domains with multiple entities are particularly difficult due to their combinatorial complexity. GCRL facilitates generalization across goals and the use of subgoal structure, but struggles with high-dimensional observations and combinatorial state-spaces, especially under sparse reward. We employ a two-level hierarchy composed of a value-based GCRL agent and a factored subgoal-generating conditional diffusion model. The RL agent and subgoal generator are trained independently and composed post hoc through selective subgoal generation based on the value function, making the approach modular and compatible with existing GCRL algorithms. We introduce new variations to benchmark tasks that highlight the challenges of multi-entity domains, and show that our method consistently boosts performance of the underlying RL agent on image-based long-horizon tasks with sparse rewards, achieving over 150% higher success rates on the hardest task in our suite and generalizing to increasing horizons and numbers of entities. Rollout videos are provided at: https://sites.google.com/view/hecrl

Zohar Rimon, Elisei Shafer, Tal Tepper et al.

Palpation, the use of touch in medical examination, is almost exclusively performed by humans. We investigate a proof of concept for an artificial palpation method based on self-supervised learning. Our key idea is that an encoder-decoder framework can learn a $\textit{representation}$ from a sequence of tactile measurements that contains all the relevant information about the palpated object. We conjecture that such a representation can be used for downstream tasks such as tactile imaging and change detection. With enough training data, it should capture intricate patterns in the tactile measurements that go beyond a simple map of forces -- the current state of the art. To validate our approach, we both develop a simulation environment and collect a real-world dataset of soft objects and corresponding ground truth images obtained by magnetic resonance imaging (MRI). We collect palpation sequences using a robot equipped with a tactile sensor, and train a model that predicts sensory readings at different positions on the object. We investigate the representation learned in this process, and demonstrate its use in imaging and change detection.

Ev Zisselman, Mirco Mutti, Shelly Francis-Meretzki et al.

Behavioral cloning is a simple yet effective technique for learning sequential decision-making from demonstrations. Recently, it has gained prominence as the core of foundation models for the physical world, where achieving generalization requires countless demonstrations of a multitude of tasks. Typically, a human expert with full information on the task demonstrates a (nearly) optimal behavior. In this paper, we propose to hide some of the task's information from the demonstrator. This ``blindfolded'' expert is compelled to employ non-trivial exploration to solve the task. We show that cloning the blindfolded expert generalizes better to unseen tasks than its fully-informed counterpart. We conduct experiments of real-world robot peg insertion tasks with (limited) human demonstrations, alongside videogames from the Procgen benchmark. Additionally, we support our findings with theoretical analysis, which confirms that the generalization error scales with $\sqrt{I/m}$, where $I$ measures the amount of task information available to the demonstrator, and $m$ is the number of demonstrated tasks. Both theory and practice indicate that cloning blindfolded experts generalizes better with fewer demonstrated tasks. Project page with videos and code: https://sites.google.com/view/blindfoldedexperts/home

Mirco Mutti, Jeongyeol Kwon, Shie Mannor et al.

Contextual multi-armed bandits are a popular choice to model sequential decision-making. E.g., in a healthcare application we may perform various tests to asses a patient condition (exploration) and then decide on the best treatment to give (exploitation). When humans design strategies, they aim for the exploration to be fast, since the patient's health is at stake, and easy to interpret for a physician overseeing the process. However, common bandit algorithms are nothing like that: The regret caused by exploration scales with $\sqrt{H}$ over $H$ rounds and decision strategies are based on opaque statistical considerations. In this paper, we use an original classification view to meta learn interpretable and fast exploration plans for a fixed collection of bandits $\mathbb{M}$. The plan is prescribed by an interpretable decision tree probing decisions' payoff to classify the test bandit. The test regret of the plan in the stochastic and contextual setting scales with $O (λ^{-2} C_λ (\mathbb{M}) \log^2 (MH))$, being $M$ the size of $\mathbb{M}$, $λ$ a separation parameter over the bandits, and $C_λ(\mathbb{M})$ a novel classification-coefficient that fundamentally links meta learning bandits with classification. Through a nearly matching lower bound, we show that $C_λ(\mathbb{M})$ inherently captures the complexity of the setting.

Mirco Mutti, Aviv Tamar

Meta reinforcement learning sets a distribution over a set of tasks on which the agent can train at will, then is asked to learn an optimal policy for any test task efficiently. In this paper, we consider a finite set of tasks modeled through Markov decision processes with various dynamics. We assume to have endured a long training phase, from which the set of tasks is perfectly recovered, and we focus on regret minimization against the optimal policy in the unknown test task. Under a separation condition that states the existence of a state-action pair revealing a task against another, Chen et al. (2022) show that $O(M^2 \log(H))$ regret can be achieved, where $M, H$ are the number of tasks in the set and test episodes, respectively. In our first contribution, we demonstrate that the latter rate is nearly optimal by developing a novel lower bound for test-time regret minimization under separation, showing that a linear dependence with $M$ is unavoidable. Then, we present a family of stronger yet reasonable assumptions beyond separation, which we call strong identifiability, enabling algorithms achieving fast rates $\log (H)$ and sublinear dependence with $M$ simultaneously. Our paper provides a new understanding of the statistical barriers of test-time regret minimization and when fast rates can be achieved.

Zohar Rimon, Tom Jurgenson, Orr Krupnik et al.

Meta-reinforcement learning (meta-RL) is a promising framework for tackling challenging domains requiring efficient exploration. Existing meta-RL algorithms are characterized by low sample efficiency, and mostly focus on low-dimensional task distributions. In parallel, model-based RL methods have been successful in solving partially observable MDPs, of which meta-RL is a special case. In this work, we leverage this success and propose a new model-based approach to meta-RL, based on elements from existing state-of-the-art model-based and meta-RL methods. We demonstrate the effectiveness of our approach on common meta-RL benchmark domains, attaining greater return with better sample efficiency (up to $15\times$) while requiring very little hyperparameter tuning. In addition, we validate our approach on a slate of more challenging, higher-dimensional domains, taking a step towards real-world generalizing agents.

Tom Jurgenson, Aviv Tamar

Recently, a simple yet effective algorithm -- goal-conditioned supervised-learning (GCSL) -- was proposed to tackle goal-conditioned reinforcement-learning. GCSL is based on the principle of hindsight learning: by observing states visited in previously executed trajectories and treating them as attained goals, GCSL learns the corresponding actions via supervised learning. However, GCSL only learns a goal-conditioned policy, discarding other information in the process. Our insight is that the same hindsight principle can be used to learn to predict goal-conditioned sub-goals from the same trajectory. Based on this idea, we propose Trajectory Iterative Learner (TraIL), an extension of GCSL that further exploits the information in a trajectory, and uses it for learning to predict both actions and sub-goals. We investigate the settings in which TraIL can make better use of the data, and discover that for several popular problem settings, replacing real goals in GCSL with predicted TraIL sub-goals allows the agent to reach a greater set of goal states using the exact same data as GCSL, thereby improving its overall performance.

Gal Leibovich, Guy Jacob, Shadi Endrawis et al.

A practical approach to learning robot skills, often termed sim2real, is to train control policies in simulation and then deploy them on a real robot. Popular techniques to improve the sim2real transfer build on domain randomization (DR) -- training the policy on a diverse set of randomly generated domains with the hope of better generalization to the real world. Due to the large number of hyper-parameters in both the policy learning and DR algorithms, one often ends up with a large number of trained policies, where choosing the best policy among them demands costly evaluation on the real robot. In this work we ask - can we rank the policies without running them in the real world? Our main idea is that a predefined set of real world data can be used to evaluate all policies, using out-of-distribution detection (OOD) techniques. In a sense, this approach can be seen as a `unit test' to evaluate policies before any real world execution. However, we find that by itself, the OOD score can be inaccurate and very sensitive to the particular OOD method. Our main contribution is a simple-yet-effective policy score that combines OOD with an evaluation in simulation. We show that our score - VSDR - can significantly improve the accuracy of policy ranking without requiring additional real world data. We evaluate the effectiveness of VSDR on sim2real transfer in a robotic grasping task with image inputs. We extensively evaluate different DR parameters and OOD methods, and show that VSDR improves policy selection across the board. More importantly, our method achieves significantly better ranking, and uses significantly less data compared to baselines. Project website is available at https://sites.google.com/view/vsdr/home.

Aviv Tamar, Daniel Soudry, Ev Zisselman

In the Bayesian reinforcement learning (RL) setting, a prior distribution over the unknown problem parameters -- the rewards and transitions -- is assumed, and a policy that optimizes the (posterior) expected return is sought. A common approximation, which has been recently popularized as meta-RL, is to train the agent on a sample of $N$ problem instances from the prior, with the hope that for large enough $N$, good generalization behavior to an unseen test instance will be obtained. In this work, we study generalization in Bayesian RL under the probably approximately correct (PAC) framework, using the method of algorithmic stability. Our main contribution is showing that by adding regularization, the optimal policy becomes stable in an appropriate sense. Most stability results in the literature build on strong convexity of the regularized loss -- an approach that is not suitable for RL as Markov decision processes (MDPs) are not convex. Instead, building on recent results of fast convergence rates for mirror descent in regularized MDPs, we show that regularized MDPs satisfy a certain quadratic growth criterion, which is sufficient to establish stability. This result, which may be of independent interest, allows us to study the effect of regularization on generalization in the Bayesian RL setting.

Shadi Endrawis, Gal Leibovich, Guy Jacob et al.

A practical approach to robot reinforcement learning is to first collect a large batch of real or simulated robot interaction data, using some data collection policy, and then learn from this data to perform various tasks, using offline learning algorithms. Previous work focused on manually designing the data collection policy, and on tasks where suitable policies can easily be designed, such as random picking policies for collecting data about object grasping. For more complex tasks, however, it may be difficult to find a data collection policy that explores the environment effectively, and produces data that is diverse enough for the downstream task. In this work, we propose that data collection policies should actively explore the environment to collect diverse data. In particular, we develop a simple-yet-effective goal-conditioned reinforcement-learning method that actively focuses data collection on novel observations, thereby collecting a diverse data-set. We evaluate our method on simulated robot manipulation tasks with visual inputs and show that the improved diversity of active data collection leads to significant improvements in the downstream learning tasks.

Carmel Rabinovitz, Niko Grupen, Aviv Tamar

Robotic tasks such as manipulation with visual inputs require image features that capture the physical properties of the scene, e.g., the position and configuration of objects. Recently, it has been suggested to learn such features in an unsupervised manner from simulated, self-supervised, robot interaction; the idea being that high-level physical properties are well captured by modern physical simulators, and their representation from visual inputs may transfer well to the real world. In particular, learning methods based on noise contrastive estimation have shown promising results. To robustify the simulation-to-real transfer, domain randomization (DR) was suggested for learning features that are invariant to irrelevant visual properties such as textures or lighting. In this work, however, we show that a naive application of DR to unsupervised learning based on contrastive estimation does not promote invariance, as the loss function maximizes mutual information between the features and both the relevant and irrelevant visual properties. We propose a simple modification of the contrastive loss to fix this, exploiting the fact that we can control the simulated randomization of visual properties. Our approach learns physical features that are significantly more robust to visual domain variation, as we demonstrate using both rigid and non-rigid objects.

Tal Daniel, Aviv Tamar

The recently introduced introspective variational autoencoder (IntroVAE) exhibits outstanding image generations, and allows for amortized inference using an image encoder. The main idea in IntroVAE is to train a VAE adversarially, using the VAE encoder to discriminate between generated and real data samples. However, the original IntroVAE loss function relied on a particular hinge-loss formulation that is very hard to stabilize in practice, and its theoretical convergence analysis ignored important terms in the loss. In this work, we take a step towards better understanding of the IntroVAE model, its practical implementation, and its applications. We propose the Soft-IntroVAE, a modified IntroVAE that replaces the hinge-loss terms with a smooth exponential loss on generated samples. This change significantly improves training stability, and also enables theoretical analysis of the complete algorithm. Interestingly, we show that the IntroVAE converges to a distribution that minimizes a sum of KL distance from the data distribution and an entropy term. We discuss the implications of this result, and demonstrate that it induces competitive image generation and reconstruction. Finally, we describe two applications of Soft-IntroVAE to unsupervised image translation and out-of-distribution detection, and demonstrate compelling results. Code and additional information is available on the project website -- https://taldatech.github.io/soft-intro-vae-web

Noga H. Rotman, Michael Schapira, Aviv Tamar

Recently, deep learning has been successfully applied to a variety of networking problems. A fundamental challenge is that when the operational environment for a learning-augmented system differs from its training environment, such systems often make badly informed decisions, leading to bad performance. We argue that safely deploying learning-driven systems requires being able to determine, in real time, whether system behavior is coherent, for the purpose of defaulting to a reasonable heuristic when this is not so. We term this the online safety assurance problem (OSAP). We present three approaches to quantifying decision uncertainty that differ in terms of the signal used to infer uncertainty. We illustrate the usefulness of online safety assurance in the context of the proposed deep reinforcement learning (RL) approach to video streaming. While deep RL for video streaming bests other approaches when the operational and training environments match, it is dominated by simple heuristics when the two differ. Our preliminary findings suggest that transitioning to a default policy when decision uncertainty is detected is key to enjoying the performance benefits afforded by leveraging ML without compromising on safety.

Tzvika Geft, Aviv Tamar, Ken Goldberg et al.

To compute robust 2D assembly plans, we present an approach that combines geometric planning with a deep neural network. We train the network using the Box2D physics simulator with added stochastic noise to yield robustness scores--the success probabilities of planned assembly motions. As running a simulation for every assembly motion is impractical, we train a convolutional neural network to map assembly operations, given as an image pair of the subassemblies before and after they are mated, to a robustness score. The neural network prediction is used within a planner to quickly prune out motions that are not robust. We demonstrate this approach on two-handed planar assemblies, where the motions are one-step translations. Results suggest that the neural network can learn robustness to plan robust sequences an order of magnitude faster than physics simulation.

Ron Dorfman, Idan Shenfeld, Aviv Tamar

Consider the following instance of the Offline Meta Reinforcement Learning (OMRL) problem: given the complete training logs of $N$ conventional RL agents, trained on $N$ different tasks, design a meta-agent that can quickly maximize reward in a new, unseen task from the same task distribution. In particular, while each conventional RL agent explored and exploited its own different task, the meta-agent must identify regularities in the data that lead to effective exploration/exploitation in the unseen task. Here, we take a Bayesian RL (BRL) view, and seek to learn a Bayes-optimal policy from the offline data. Building on the recent VariBAD BRL approach, we develop an off-policy BRL method that learns to plan an exploration strategy based on an adaptive neural belief estimate. However, learning to infer such a belief from offline data brings a new identifiability issue we term MDP ambiguity. We characterize the problem, and suggest resolutions via data collection and modification procedures. Finally, we evaluate our framework on a diverse set of domains, including difficult sparse reward tasks, and demonstrate learning of effective exploration behavior that is qualitatively different from the exploration used by any RL agent in the data.

Roi Bar-Zur, Ittay Eyal, Aviv Tamar

A proof of work (PoW) blockchain protocol distributes rewards to its participants, called miners, according to their share of the total computational power. Sufficiently large miners can perform selfish mining - deviate from the protocol to gain more than their fair share. Such systems are thus secure if all miners are smaller than a threshold size so their best response is following the protocol. To find the threshold, one has to identify the optimal strategy for miners of different sizes, i.e., solve a Markov Decision Process (MDP). However, because of the PoW difficulty adjustment mechanism, the miners' utility is a non-linear ratio function. We therefore call this an Average Reward Ratio (ARR) MDP. Sapirshtein et al.\ were the first to solve ARR MDPs by solving a series of standard MDPs that converge to the ARR MDP solution. In this work, we present a novel technique for solving an ARR MDP by solving a single standard MDP. The crux of our approach is to augment the MDP such that it terminates randomly, within an expected number of rounds. We call this Probabilistic Termination Optimization (PTO), and the technique applies to any MDP whose utility is a ratio function. We bound the approximation error of PTO - it is inversely proportional to the expected number of rounds before termination, a parameter that we control. Empirically, PTO's complexity is an order of magnitude lower than the state of the art. PTO can be easily applied to different blockchains. We use it to tighten the bound on the threshold for selfish mining in Ethereum.

Kara Liu, Thanard Kurutach, Christine Tung et al.

In visual planning (VP), an agent learns to plan goal-directed behavior from observations of a dynamical system obtained offline, e.g., images obtained from self-supervised robot interaction. Most previous works on VP approached the problem by planning in a learned latent space, resulting in low-quality visual plans, and difficult training algorithms. Here, instead, we propose a simple VP method that plans directly in image space and displays competitive performance. We build on the semi-parametric topological memory (SPTM) method: image samples are treated as nodes in a graph, the graph connectivity is learned from image sequence data, and planning can be performed using conventional graph search methods. We propose two modifications on SPTM. First, we train an energy-based graph connectivity function using contrastive predictive coding that admits stable training. Second, to allow zero-shot planning in new domains, we learn a conditional VAE model that generates images given a context of the domain, and use these hallucinated samples for building the connectivity graph and planning. We show that this simple approach significantly outperform the state-of-the-art VP methods, in terms of both plan interpretability and success rate when using the plan to guide a trajectory-following controller. Interestingly, our method can pick up non-trivial visual properties of objects, such as their geometry, and account for it in the plans.

Tom Jurgenson, Or Avner, Edward Groshev et al.

Many AI problems, in robotics and other domains, are goal-based, essentially seeking trajectories leading to various goal states. Reinforcement learning (RL), building on Bellman's optimality equation, naturally optimizes for a single goal, yet can be made multi-goal by augmenting the state with the goal. Instead, we propose a new RL framework, derived from a dynamic programming equation for the all pairs shortest path (APSP) problem, which naturally solves multi-goal queries. We show that this approach has computational benefits for both standard and approximate dynamic programming. Interestingly, our formulation prescribes a novel protocol for computing a trajectory: instead of predicting the next state given its predecessor, as in standard RL, a goal-conditioned trajectory is constructed by first predicting an intermediate state between start and goal, partitioning the trajectory into two. Then, recursively, predicting intermediate points on each sub-segment, until a complete trajectory is obtained. We call this trajectory structure a sub-goal tree. Building on it, we additionally extend the policy gradient methodology to recursively predict sub-goals, resulting in novel goal-based algorithms. Finally, we apply our method to neural motion planning, where we demonstrate significant improvements compared to standard RL on navigating a 7-DoF robot arm between obstacles.

Ev Zisselman, Aviv Tamar

The effective application of neural networks in the real-world relies on proficiently detecting out-of-distribution examples. Contemporary methods seek to model the distribution of feature activations in the training data for adequately distinguishing abnormalities, and the state-of-the-art method uses Gaussian distribution models. In this work, we present a novel approach that improves upon the state-of-the-art by leveraging an expressive density model based on normalizing flows. We introduce the residual flow, a novel flow architecture that learns the residual distribution from a base Gaussian distribution. Our model is general, and can be applied to any data that is approximately Gaussian. For out of distribution detection in image datasets, our approach provides a principled improvement over the state-of-the-art. Specifically, we demonstrate the effectiveness of our method in ResNet and DenseNet architectures trained on various image datasets. For example, on a ResNet trained on CIFAR-100 and evaluated on detection of out-of-distribution samples from the ImageNet dataset, holding the true positive rate (TPR) at $95\%$, we improve the true negative rate (TNR) from $56.7\%$ (current state-of-the-art) to $77.5\%$ (ours).

Tal Daniel, Thanard Kurutach, Aviv Tamar

In anomaly detection (AD), one seeks to identify whether a test sample is abnormal, given a data set of normal samples. A recent and promising approach to AD relies on deep generative models, such as variational autoencoders (VAEs), for unsupervised learning of the normal data distribution. In semi-supervised AD (SSAD), the data also includes a small sample of labeled anomalies. In this work, we propose two variational methods for training VAEs for SSAD. The intuitive idea in both methods is to train the encoder to `separate' between latent vectors for normal and outlier data. We show that this idea can be derived from principled probabilistic formulations of the problem, and propose simple and effective algorithms. Our methods can be applied to various data types, as we demonstrate on SSAD datasets ranging from natural images to astronomy and medicine, can be combined with any VAE model architecture, and are naturally compatible with ensembling. When comparing to state-of-the-art SSAD methods that are not specific to particular data types, we obtain marked improvement in outlier detection.

Yi Wu, Yuxin Wu, Aviv Tamar et al.

We introduce a new memory architecture, Bayesian Relational Memory (BRM), to improve the generalization ability for semantic visual navigation agents in unseen environments, where an agent is given a semantic target to navigate towards. BRM takes the form of a probabilistic relation graph over semantic entities (e.g., room types), which allows (1) capturing the layout prior from training environments, i.e., prior knowledge, (2) estimating posterior layout at test time, i.e., memory update, and (3) efficient planning for navigation, altogether. We develop a BRM agent consisting of a BRM module for producing sub-goals and a goal-conditioned locomotion module for control. When testing in unseen environments, the BRM agent outperforms baselines that do not explicitly utilize the probabilistic relational memory structure

Tom Jurgenson, Edward Groshev, Aviv Tamar

Many AI problems, in robotics and other domains, are goal-directed, essentially seeking a trajectory leading to some goal state. In such problems, the way we choose to represent a trajectory underlies algorithms for trajectory prediction and optimization. Interestingly, most all prior work in imitation and reinforcement learning builds on a sequential trajectory representation -- calculating the next state in the trajectory given its predecessors. We propose a different perspective: a goal-conditioned trajectory can be represented by first selecting an intermediate state between start and goal, partitioning the trajectory into two. Then, recursively, predicting intermediate points on each sub-segment, until a complete trajectory is obtained. We call this representation a sub-goal tree, and building on it, we develop new methods for trajectory prediction, learning, and optimization. We show that in a supervised learning setting, sub-goal trees better account for trajectory variability, and can predict trajectories exponentially faster at test time by leveraging a concurrent computation. Then, for optimization, we derive a new dynamic programming equation for sub-goal trees, and use it to develop new planning and reinforcement learning algorithms. These algorithms, which are not based on the standard Bellman equation, naturally account for hierarchical sub-goal structure in a task. Empirical results on motion planning domains show that the sub-goal tree framework significantly improves both accuracy and prediction time.

Tom Jurgenson, Aviv Tamar

Motion planning is an essential component in most of today's robotic applications. In this work, we consider the learning setting, where a set of solved motion planning problems is used to improve the efficiency of motion planning on different, yet similar problems. This setting is important in applications with rapidly changing environments such as in e-commerce, among others. We investigate a general deep learning based approach, where a neural network is trained to map an image of the domain, the current robot state, and a goal robot state to the next robot state in the plan. We focus on the learning algorithm, and compare supervised learning methods with reinforcement learning (RL) algorithms. We first establish that supervised learning approaches are inferior in their accuracy due to insufficient data on the boundary of the obstacles, an issue that RL methods mitigate by actively exploring the domain. We then propose a modification of the popular DDPG RL algorithm that is tailored to motion planning domains, by exploiting the known model in the problem and the set of solved plans in the data. We show that our algorithm, dubbed DDPG-MP, significantly improves the accuracy of the learned motion planning policy. Finally, we show that given enough training data, our method can plan significantly faster on novel domains than off-the-shelf sampling based motion planners. Results of our experiments are shown in https://youtu.be/wHQ4Y4mBRb8.

Angelina Wang, Thanard Kurutach, Kara Liu et al.

Planning for robotic manipulation requires reasoning about the changes a robot can affect on objects. When such interactions can be modelled analytically, as in domains with rigid objects, efficient planning algorithms exist. However, in both domestic and industrial domains, the objects of interest can be soft, or deformable, and hard to model analytically. For such cases, we posit that a data-driven modelling approach is more suitable. In recent years, progress in deep generative models has produced methods that learn to `imagine' plausible images from data. Building on the recent Causal InfoGAN generative model, in this work we learn to imagine goal-directed object manipulation directly from raw image data of self-supervised interaction of the robot with the object. After learning, given a goal observation of the system, our model can generate an imagined plan -- a sequence of images that transition the object into the desired goal. To execute the plan, we use it as a reference trajectory to track with a visual servoing controller, which we also learn from the data as an inverse dynamics model. In a simulated manipulation task, we show that separating the problem into visual planning and visual tracking control is more sample efficient and more interpretable than alternative data-driven approaches. We further demonstrate our approach on learning to imagine and execute in 3 environments, the final of which is deformable rope manipulation on a PR2 robot.

Xinyi Ren, Jianlan Luo, Eugen Solowjow et al.

Accurate state estimation is a fundamental component of robotic control. In robotic manipulation tasks, as is our focus in this work, state estimation is essential for identifying the positions of objects in the scene, forming the basis of the manipulation plan. However, pose estimation typically requires expensive 3D cameras or additional instrumentation such as fiducial markers to perform accurately. Recently, Tobin et al.~introduced an approach to pose estimation based on domain randomization, where a neural network is trained to predict pose directly from a 2D image of the scene. The network is trained on computer-generated images with a high variation in textures and lighting, thereby generalizing to real-world images. In this work, we investigate how to improve the accuracy of domain randomization based pose estimation. Our main idea is that active perception -- moving the robot to get a better estimate of pose -- can be trained in simulation and transferred to real using domain randomization. In our approach, the robot trains in a domain-randomized simulation how to estimate pose from a \emph{sequence} of images. We show that our approach can significantly improve the accuracy of standard pose estimation in several scenarios: when the robot holding an object moves, when reference objects are moved in the scene, or when the camera is moved around the object.

Margaret P. Chapman, Jonathan Lacotte, Aviv Tamar et al.

A classic reachability problem for safety of dynamic systems is to compute the set of initial states from which the state trajectory is guaranteed to stay inside a given constraint set over a given time horizon. In this paper, we leverage existing theory of reachability analysis and risk measures to devise a risk-sensitive reachability approach for safety of stochastic dynamic systems under non-adversarial disturbances over a finite time horizon. Specifically, we first introduce the notion of a risk-sensitive safe set as a set of initial states from which the risk of large constraint violations can be reduced to a required level via a control policy, where risk is quantified using the Conditional Value-at-Risk (CVaR) measure. Second, we show how the computation of a risk-sensitive safe set can be reduced to the solution to a Markov Decision Process (MDP), where cost is assessed according to CVaR. Third, leveraging this reduction, we devise a tractable algorithm to approximate a risk-sensitive safe set, and provide theoretical arguments about its correctness. Finally, we present a realistic example inspired from stormwater catchment design to demonstrate the utility of risk-sensitive reachability analysis. In particular, our approach allows a practitioner to tune the level of risk sensitivity from worst-case (which is typical for Hamilton-Jacobi reachability analysis) to risk-neutral (which is the case for stochastic reachability analysis).

Orr Krupnik, Igor Mordatch, Aviv Tamar

We consider model-based reinforcement learning (MBRL) in 2-agent, high-fidelity continuous control problems -- an important domain for robots interacting with other agents in the same workspace. For non-trivial dynamical systems, MBRL typically suffers from accumulating errors. Several recent studies have addressed this problem by learning latent variable models for trajectory segments and optimizing over behavior in the latent space. In this work, we investigate whether this approach can be extended to 2-agent competitive and cooperative settings. The fundamental challenge is how to learn models that capture interactions between agents, yet are disentangled to allow for optimization of each agent behavior separately. We propose such models based on a disentangled variational auto-encoder, and demonstrate our approach on a simulated 2-robot manipulation task, where one robot can either help or distract the other. We show that our approach has better sample efficiency than a strong model-free RL baseline, and can learn both cooperative and adversarial behavior from the same data.

Yi Wu, Yuxin Wu, Aviv Tamar et al.

Building deep reinforcement learning agents that can generalize and adapt to unseen environments remains a fundamental challenge for AI. This paper describes progresses on this challenge in the context of man-made environments, which are visually diverse but contain intrinsic semantic regularities. We propose a hybrid model-based and model-free approach, LEArning and Planning with Semantics (LEAPS), consisting of a multi-target sub-policy that acts on visual inputs, and a Bayesian model over semantic structures. When placed in an unseen environment, the agent plans with the semantic model to make high-level decisions, proposes the next sub-target for the sub-policy to execute, and updates the semantic model based on new observations. We perform experiments in visual navigation tasks using House3D, a 3D environment that contains diverse human-designed indoor scenes with real-world objects. LEAPS outperforms strong baselines that do not explicitly plan using the semantic content.

Dror Freirich, Ron Meir, Aviv Tamar

The recently proposed distributional approach to reinforcement learning (DiRL) is centered on learning the distribution of the reward-to-go, often referred to as the value distribution. In this work, we show that the distributional Bellman equation, which drives DiRL methods, is equivalent to a generative adversarial network (GAN) model. In this formulation, DiRL can be seen as learning a deep generative model of the value distribution, driven by the discrepancy between the distribution of the current value, and the distribution of the sum of current reward and next value. We use this insight to propose a GAN-based approach to DiRL, which leverages the strengths of GANs in learning distributions of high-dimensional data. In particular, we show that our GAN approach can be used for DiRL with multivariate rewards, an important setting which cannot be tackled with prior methods. The multivariate setting also allows us to unify learning the distribution of values and state transitions, and we exploit this idea to devise a novel exploration method that is driven by the discrepancy in estimating both values and states.

Thanard Kurutach, Aviv Tamar, Ge Yang et al.

In recent years, deep generative models have been shown to 'imagine' convincing high-dimensional observations such as images, audio, and even video, learning directly from raw data. In this work, we ask how to imagine goal-directed visual plans -- a plausible sequence of observations that transition a dynamical system from its current configuration to a desired goal state, which can later be used as a reference trajectory for control. We focus on systems with high-dimensional observations, such as images, and propose an approach that naturally combines representation learning and planning. Our framework learns a generative model of sequential observations, where the generative process is induced by a transition in a low-dimensional planning model, and an additional noise. By maximizing the mutual information between the generated observations and the transition in the planning model, we obtain a low-dimensional representation that best explains the causal nature of the data. We structure the planning model to be compatible with efficient planning algorithms, and we propose several such models based on either discrete or continuous states. Finally, to generate a visual plan, we project the current and goal observations onto their respective states in the planning model, plan a trajectory, and then use the generative model to transform the trajectory to a sequence of observations. We demonstrate our method on imagining plausible visual plans of rope manipulation.

Garrett Thomas, Melissa Chien, Aviv Tamar et al.

In this work, motivated by recent manufacturing trends, we investigate autonomous robotic assembly. Industrial assembly tasks require contact-rich manipulation skills, which are challenging to acquire using classical control and motion planning approaches. Consequently, robot controllers for assembly domains are presently engineered to solve a particular task, and cannot easily handle variations in the product or environment. Reinforcement learning (RL) is a promising approach for autonomously acquiring robot skills that involve contact-rich dynamics. However, RL relies on random exploration for learning a control policy, which requires many robot executions, and often gets trapped in locally suboptimal solutions. Instead, we posit that prior knowledge, when available, can improve RL performance. We exploit the fact that in modern assembly domains, geometric information about the task is readily available via the CAD design files. We propose to leverage this prior knowledge by guiding RL along a geometric motion plan, calculated using the CAD data. We show that our approach effectively improves over traditional control approaches for tracking the motion plan, and can solve assembly tasks that require high precision, even without accurate state estimation. In addition, we propose a neural network architecture that can learn to track the motion plan, and generalize the assembly controller to changes in the object positions.

Thanard Kurutach, Ignasi Clavera, Yan Duan et al.

Model-free reinforcement learning (RL) methods are succeeding in a growing number of tasks, aided by recent advances in deep learning. However, they tend to suffer from high sample complexity, which hinders their use in real-world domains. Alternatively, model-based reinforcement learning promises to reduce sample complexity, but tends to require careful tuning and to date have succeeded mainly in restrictive domains where simple models are sufficient for learning. In this paper, we analyze the behavior of vanilla model-based reinforcement learning methods when deep neural networks are used to learn both the model and the policy, and show that the learned policy tends to exploit regions where insufficient data is available for the model to be learned, causing instability in training. To overcome this issue, we propose to use an ensemble of models to maintain the model uncertainty and regularize the learning process. We further show that the use of likelihood ratio derivatives yields much more stable learning than backpropagation through time. Altogether, our approach Model-Ensemble Trust-Region Policy Optimization (ME-TRPO) significantly reduces the sample complexity compared to model-free deep RL methods on challenging continuous control benchmark tasks.