Sandy H. Huang

Tuomas Haarnoja, Ben Moran, Guy Lever et al. · deepmind

We investigate whether Deep Reinforcement Learning (Deep RL) is able to synthesize sophisticated and safe movement skills for a low-cost, miniature humanoid robot that can be composed into complex behavioral strategies in dynamic environments. We used Deep RL to train a humanoid robot with 20 actuated joints to play a simplified one-versus-one (1v1) soccer game. The resulting agent exhibits robust and dynamic movement skills such as rapid fall recovery, walking, turning, kicking and more; and it transitions between them in a smooth, stable, and efficient manner. The agent's locomotion and tactical behavior adapts to specific game contexts in a way that would be impractical to manually design. The agent also developed a basic strategic understanding of the game, and learned, for instance, to anticipate ball movements and to block opponent shots. Our agent was trained in simulation and transferred to real robots zero-shot. We found that a combination of sufficiently high-frequency control, targeted dynamics randomization, and perturbations during training in simulation enabled good-quality transfer. Although the robots are inherently fragile, basic regularization of the behavior during training led the robots to learn safe and effective movements while still performing in a dynamic and agile way -- well beyond what is intuitively expected from the robot. Indeed, in experiments, they walked 181% faster, turned 302% faster, took 63% less time to get up, and kicked a ball 34% faster than a scripted baseline, while efficiently combining the skills to achieve the longer term objectives.

Dhruva Tirumala, Thomas Lampe, Jose Enrique Chen et al. · deepmind

Replaying data is a principal mechanism underlying the stability and data efficiency of off-policy reinforcement learning (RL). We present an effective yet simple framework to extend the use of replays across multiple experiments, minimally adapting the RL workflow for sizeable improvements in controller performance and research iteration times. At its core, Replay Across Experiments (RaE) involves reusing experience from previous experiments to improve exploration and bootstrap learning while reducing required changes to a minimum in comparison to prior work. We empirically show benefits across a number of RL algorithms and challenging control domains spanning both locomotion and manipulation, including hard exploration tasks from egocentric vision. Through comprehensive ablations, we demonstrate robustness to the quality and amount of data available and various hyperparameter choices. Finally, we discuss how our approach can be applied more broadly across research life cycles and can increase resilience by reloading data across random seeds or hyperparameter variations.

Markus Wulfmeier, Michael Bloesch, Nino Vieillard et al. · deepmind

The majority of language model training builds on imitation learning. It covers pretraining, supervised fine-tuning, and affects the starting conditions for reinforcement learning from human feedback (RLHF). The simplicity and scalability of maximum likelihood estimation (MLE) for next token prediction led to its role as predominant paradigm. However, the broader field of imitation learning can more effectively utilize the sequential structure underlying autoregressive generation. We focus on investigating the inverse reinforcement learning (IRL) perspective to imitation, extracting rewards and directly optimizing sequences instead of individual token likelihoods and evaluate its benefits for fine-tuning large language models. We provide a new angle, reformulating inverse soft-Q-learning as a temporal difference regularized extension of MLE. This creates a principled connection between MLE and IRL and allows trading off added complexity with increased performance and diversity of generations in the supervised fine-tuning (SFT) setting. We find clear advantages for IRL-based imitation, in particular for retaining diversity while maximizing task performance, rendering IRL a strong alternative on fixed SFT datasets even without online data generation. Our analysis of IRL-extracted reward functions further indicates benefits for more robust reward functions via tighter integration of supervised and preference-based LLM post-training.

Joe Watson, Sandy H. Huang, Nicolas Heess

Imitation learning methods seek to learn from an expert either through behavioral cloning (BC) of the policy or inverse reinforcement learning (IRL) of the reward. Such methods enable agents to learn complex tasks from humans that are difficult to capture with hand-designed reward functions. Choosing BC or IRL for imitation depends on the quality and state-action coverage of the demonstrations, as well as additional access to the Markov decision process. Hybrid strategies that combine BC and IRL are not common, as initial policy optimization against inaccurate rewards diminishes the benefit of pretraining the policy with BC. This work derives an imitation method that captures the strengths of both BC and IRL. In the entropy-regularized ('soft') reinforcement learning setting, we show that the behaviour-cloned policy can be used as both a shaped reward and a critic hypothesis space by inverting the regularized policy update. This coherency facilitates fine-tuning cloned policies using the reward estimate and additional interactions with the environment. This approach conveniently achieves imitation learning through initial behaviour cloning, followed by refinement via RL with online or offline data sources. The simplicity of the approach enables graceful scaling to high-dimensional and vision-based tasks, with stable learning and minimal hyperparameter tuning, in contrast to adversarial approaches. For the open-source implementation and simulation results, see https://joemwatson.github.io/csil/.

Dhruva Tirumala, Markus Wulfmeier, Ben Moran et al. · deepmind

We apply multi-agent deep reinforcement learning (RL) to train end-to-end robot soccer policies with fully onboard computation and sensing via egocentric RGB vision. This setting reflects many challenges of real-world robotics, including active perception, agile full-body control, and long-horizon planning in a dynamic, partially-observable, multi-agent domain. We rely on large-scale, simulation-based data generation to obtain complex behaviors from egocentric vision which can be successfully transferred to physical robots using low-cost sensors. To achieve adequate visual realism, our simulation combines rigid-body physics with learned, realistic rendering via multiple Neural Radiance Fields (NeRFs). We combine teacher-based multi-agent RL and cross-experiment data reuse to enable the discovery of sophisticated soccer strategies. We analyze active-perception behaviors including object tracking and ball seeking that emerge when simply optimizing perception-agnostic soccer play. The agents display equivalent levels of performance and agility as policies with access to privileged, ground-truth state. To our knowledge, this paper constitutes a first demonstration of end-to-end training for multi-agent robot soccer, mapping raw pixel observations to joint-level actions, that can be deployed in the real world. Videos of the game-play and analyses can be seen on our website https://sites.google.com/view/vision-soccer .

Jonas Buchli, Brendan Tracey, Tomislav Andric et al. · deepmind

Improved low-frequency sensitivity of gravitational wave observatories would unlock study of intermediate-mass black hole mergers, binary black hole eccentricity, and provide early warnings for multi-messenger observations of binary neutron star mergers. Today's mirror stabilization control injects harmful noise, constituting a major obstacle to sensitivity improvements. We eliminated this noise through Deep Loop Shaping, a reinforcement learning method using frequency domain rewards. We proved our methodology on the LIGO Livingston Observatory (LLO). Our controller reduced control noise in the 10--30Hz band by over 30x, and up to 100x in sub-bands surpassing the design goal motivated by the quantum limit. These results highlight the potential of Deep Loop Shaping to improve current and future GW observatories, and more broadly instrumentation and control systems.

Annie S. Chen, Philemon Brakel, Antonia Bronars et al. · deepmind

Learning-based methods for dexterous manipulation have made notable progress in recent years. However, learned policies often still lack reliability and exhibit limited robustness to important factors of variation. One failure pattern that can be observed across many settings is that policies idle, i.e. they cease to move beyond a small region of states when they reach certain states. This policy idling is often a reflection of the training data. For instance, it can occur when the data contains small actions in areas where the robot needs to perform high-precision motions, e.g., when preparing to grasp an object or object insertion. Prior works have tried to mitigate this phenomenon e.g. by filtering the training data or modifying the control frequency. However, these approaches can negatively impact policy performance in other ways. As an alternative, we investigate how to leverage the detectability of idling behavior to inform exploration and policy improvement. Our approach, Pause-Induced Perturbations (PIP), applies perturbations at detected idling states, thus helping it to escape problematic basins of attraction. On a range of challenging simulated dual-arm tasks, we find that this simple approach can already noticeably improve test-time performance, with no additional supervision or training. Furthermore, since the robot tends to idle at critical points in a movement, we also find that learning from the resulting episodes leads to better iterative policy improvement compared to prior approaches. Our perturbation strategy also leads to a 15-35% improvement in absolute success rate on a real-world insertion task that requires complex multi-finger manipulation.

Eliza Kosoy, Adrian Liu, Jasmine Collins et al.

Despite recent progress in reinforcement learning (RL), RL algorithms for exploration still remain an active area of research. Existing methods often focus on state-based metrics, which do not consider the underlying causal structures of the environment, and while recent research has begun to explore RL environments for causal learning, these environments primarily leverage causal information through causal inference or induction rather than exploration. In contrast, human children - some of the most proficient explorers - have been shown to use causal information to great benefit. In this work, we introduce a novel RL environment designed with a controllable causal structure, which allows us to evaluate exploration strategies used by both agents and children in a unified environment. In addition, through experimentation on both computation models and children, we demonstrate that there are significant differences between information-gain optimal RL exploration in causal environments and the exploration of children in the same environments. We conclude with a discussion of how these findings may inspire new directions of research into efficient exploration and disambiguation of causal structures for RL algorithms.

Abbas Abdolmaleki, Sandy H. Huang, Giulia Vezzani et al.

Many advances that have improved the robustness and efficiency of deep reinforcement learning (RL) algorithms can, in one way or another, be understood as introducing additional objectives or constraints in the policy optimization step. This includes ideas as far ranging as exploration bonuses, entropy regularization, and regularization toward teachers or data priors. Often, the task reward and auxiliary objectives are in conflict, and in this paper we argue that this makes it natural to treat these cases as instances of multi-objective (MO) optimization problems. We demonstrate how this perspective allows us to develop novel and more effective RL algorithms. In particular, we focus on offline RL and finetuning as case studies, and show that existing approaches can be understood as MO algorithms relying on linear scalarization. We hypothesize that replacing linear scalarization with a better algorithm can improve performance. We introduce Distillation of a Mixture of Experts (DiME), a new MORL algorithm that outperforms linear scalarization and can be applied to these non-standard MO problems. We demonstrate that for offline RL, DiME leads to a simple new algorithm that outperforms state-of-the-art. For finetuning, we derive new algorithms that learn to outperform the teacher policy.

Abbas Abdolmaleki, Sandy H. Huang, Leonard Hasenclever et al.

Many real-world problems require trading off multiple competing objectives. However, these objectives are often in different units and/or scales, which can make it challenging for practitioners to express numerical preferences over objectives in their native units. In this paper we propose a novel algorithm for multi-objective reinforcement learning that enables setting desired preferences for objectives in a scale-invariant way. We propose to learn an action distribution for each objective, and we use supervised learning to fit a parametric policy to a combination of these distributions. We demonstrate the effectiveness of our approach on challenging high-dimensional real and simulated robotics tasks, and show that setting different preferences in our framework allows us to trace out the space of nondominated solutions.

Eliza Kosoy, Jasmine Collins, David M. Chan et al.

Research in developmental psychology consistently shows that children explore the world thoroughly and efficiently and that this exploration allows them to learn. In turn, this early learning supports more robust generalization and intelligent behavior later in life. While much work has gone into developing methods for exploration in machine learning, artificial agents have not yet reached the high standard set by their human counterparts. In this work we propose using DeepMind Lab (Beattie et al., 2016) as a platform to directly compare child and agent behaviors and to develop new exploration techniques. We outline two ongoing experiments to demonstrate the effectiveness of a direct comparison, and outline a number of open research questions that we believe can be tested using this methodology.

Sandy H. Huang, Isabella Huang, Ravi Pandya et al.

Robots can learn preferences from human demonstrations, but their success depends on how informative these demonstrations are. Being informative is unfortunately very challenging, because during teaching, people typically get no transparency into what the robot already knows or has learned so far. In contrast, human students naturally provide a wealth of nonverbal feedback that reveals their level of understanding and engagement. In this work, we study how a robot can similarly provide feedback that is minimally disruptive, yet gives human teachers a better mental model of the robot learner, and thus enables them to teach more effectively. Our idea is that at any point, the robot can indicate what it thinks the correct next action is, shedding light on its current estimate of the human's preferences. We analyze how useful this feedback is, both in theory and with two user studies---one with a virtual character that tests the feedback itself, and one with a PR2 robot that uses gaze as the feedback mechanism. We find that feedback can be useful for improving both the quality of teaching and teachers' understanding of the robot's capability.

Sandy H. Huang, Martina Zambelli, Jackie Kay et al.

Robots must know how to be gentle when they need to interact with fragile objects, or when the robot itself is prone to wear and tear. We propose an approach that enables deep reinforcement learning to train policies that are gentle, both during exploration and task execution. In a reward-based learning environment, a natural approach involves augmenting the (task) reward with a penalty for non-gentleness, which can be defined as excessive impact force. However, augmenting with only this penalty impairs learning: policies get stuck in a local optimum which avoids all contact with the environment. Prior research has shown that combining auxiliary tasks or intrinsic rewards can be beneficial for stabilizing and accelerating learning in sparse-reward domains, and indeed we find that introducing a surprise-based intrinsic reward does avoid the no-contact failure case. However, we show that a simple dynamics-based surprise is not as effective as penalty-based surprise. Penalty-based surprise, based on predicting forceful contacts, has a further benefit: it encourages exploration which is contact-rich yet gentle. We demonstrate the effectiveness of the approach using a complex, tendon-powered robot hand with tactile sensors. Videos are available at http://sites.google.com/view/gentlemanipulation.

Ravi Pandya, Sandy H. Huang, Dylan Hadfield-Menell et al.

People frequently face challenging decision-making problems in which outcomes are uncertain or unknown. Artificial intelligence (AI) algorithms exist that can outperform humans at learning such tasks. Thus, there is an opportunity for AI agents to assist people in learning these tasks more effectively. In this work, we use a multi-armed bandit as a controlled setting in which to explore this direction. We pair humans with a selection of agents and observe how well each human-agent team performs. We find that team performance can beat both human and agent performance in isolation. Interestingly, we also find that an agent's performance in isolation does not necessarily correlate with the human-agent team's performance. A drop in agent performance can lead to a disproportionately large drop in team performance, or in some settings can even improve team performance. Pairing a human with an agent that performs slightly better than them can make them perform much better, while pairing them with an agent that performs the same can make them them perform much worse. Further, our results suggest that people have different exploration strategies and might perform better with agents that match their strategy. Overall, optimizing human-agent team performance requires going beyond optimizing agent performance, to understanding how the agent's suggestions will influence human decision-making.

Sandy H. Huang, Kush Bhatia, Pieter Abbeel et al.

In order to effectively interact with or supervise a robot, humans need to have an accurate mental model of its capabilities and how it acts. Learned neural network policies make that particularly challenging. We propose an approach for helping end-users build a mental model of such policies. Our key observation is that for most tasks, the essence of the policy is captured in a few critical states: states in which it is very important to take a certain action. Our user studies show that if the robot shows a human what its understanding of the task's critical states is, then the human can make a more informed decision about whether to deploy the policy, and if she does deploy it, when she needs to take control from it at execution time.

Sandy H. Huang, David Held, Pieter Abbeel et al.

The overarching goal of this work is to efficiently enable end-users to correctly anticipate a robot's behavior in novel situations. Since a robot's behavior is often a direct result of its underlying objective function, our insight is that end-users need to have an accurate mental model of this objective function in order to understand and predict what the robot will do. While people naturally develop such a mental model over time through observing the robot act, this familiarization process may be lengthy. Our approach reduces this time by having the robot model how people infer objectives from observed behavior, and then it selects those behaviors that are maximally informative. The problem of computing a posterior over objectives from observed behavior is known as Inverse Reinforcement Learning (IRL), and has been applied to robots learning human objectives. We consider the problem where the roles of human and robot are swapped. Our main contribution is to recognize that unlike robots, humans will not be exact in their IRL inference. We thus introduce two factors to define candidate approximate-inference models for human learning in this setting, and analyze them in a user study in the autonomous driving domain. We show that certain approximate-inference models lead to the robot generating example behaviors that better enable users to anticipate what it will do in novel situations. Our results also suggest, however, that additional research is needed in modeling how humans extrapolate from examples of robot behavior.

Sandy Huang, Nicolas Papernot, Ian Goodfellow et al.

Machine learning classifiers are known to be vulnerable to inputs maliciously constructed by adversaries to force misclassification. Such adversarial examples have been extensively studied in the context of computer vision applications. In this work, we show adversarial attacks are also effective when targeting neural network policies in reinforcement learning. Specifically, we show existing adversarial example crafting techniques can be used to significantly degrade test-time performance of trained policies. Our threat model considers adversaries capable of introducing small perturbations to the raw input of the policy. We characterize the degree of vulnerability across tasks and training algorithms, for a subclass of adversarial-example attacks in white-box and black-box settings. Regardless of the learned task or training algorithm, we observe a significant drop in performance, even with small adversarial perturbations that do not interfere with human perception. Videos are available at http://rll.berkeley.edu/adversarial.