PixL2R: Guiding Reinforcement Learning Using Natural Language by Mapping Pixels to Rewards

Reinforcement learning (RL), particularly in sparse reward settings, often requires prohibitively large numbers of interactions with the environment, thereby limiting its applicability to complex problems. To address this, several prior approaches have used natural language to guide the agent's exploration. However, these approaches typically operate on structured representations of the environment, and/or assume some structure in the natural language commands. In this work, we propose a model that directly maps pixels to rewards, given a free-form natural language description of the task, which can then be used for policy learning. Our experiments on the Meta-World robot manipulation domain show that language-based rewards significantly improves the sample efficiency of policy learning, both in sparse and dense reward settings.

Bayesian Robust Optimization for Imitation Learning

One of the main challenges in imitation learning is determining what action an agent should take when outside the state distribution of the demonstrations. Inverse reinforcement learning (IRL) can enable generalization to new states by learning a parameterized reward function, but these approaches still face uncertainty over the true reward function and corresponding optimal policy. Existing safe imitation learning approaches based on IRL deal with this uncertainty using a maxmin framework that optimizes a policy under the assumption of an adversarial reward function, whereas risk-neutral IRL approaches either optimize a policy for the mean or MAP reward function. While completely ignoring risk can lead to overly aggressive and unsafe policies, optimizing in a fully adversarial sense is also problematic as it can lead to overly conservative policies that perform poorly in practice. To provide a bridge between these two extremes, we propose Bayesian Robust Optimization for Imitation Learning (BROIL). BROIL leverages Bayesian reward function inference and a user specific risk tolerance to efficiently optimize a robust policy that balances expected return and conditional value at risk. Our empirical results show that BROIL provides a natural way to interpolate between return-maximizing and risk-minimizing behaviors and outperforms existing risk-sensitive and risk-neutral inverse reinforcement learning algorithms.

Efficiently Guiding Imitation Learning Algorithms with Human Gaze

Human gaze is known to be an intention-revealing signal in human demonstrations of tasks. In this work, we use gaze cues from human demonstrators to enhance the performance of state-of-the-art inverse reinforcement learning (IRL) and behavioral cloning (BC) algorithms. We propose a novel approach for utilizing gaze data in a computationally efficient manner --- encoding the human's attention as part of an auxiliary loss function, without adding any additional learnable parameters to those models and without requiring gaze data at test time. The auxiliary loss encourages a network to have convolutional activations in regions where the human's gaze fixated. We show how to augment any existing convolutional architecture with our auxiliary gaze loss (coverage-based gaze loss or CGL) that can guide learning toward a better reward function or policy. We show that our proposed approach improves performance of both BC and IRL methods on a variety of Atari games. We also compare against two baseline methods for utilizing gaze data with imitation learning methods. Our approach outperforms a baseline method, called gaze-modulated dropout (GMD), and is comparable to another method (AGIL) which uses gaze as input to the network and thus increases the amount of learnable parameters.

Safe Imitation Learning via Fast Bayesian Reward Inference from Preferences

Bayesian reward learning from demonstrations enables rigorous safety and uncertainty analysis when performing imitation learning. However, Bayesian reward learning methods are typically computationally intractable for complex control problems. We propose a highly efficient Bayesian reward learning algorithm that scales to high-dimensional imitation learning problems by first pre-training a low-dimensional feature encoding via self-supervised tasks and then leveraging preferences over demonstrations to perform fast Bayesian inference. We evaluate our proposed approach on the task of learning to play Atari games from demonstrations, without access to the game score. For Atari games our approach enables us to generate 100,000 samples from the posterior over reward functions in only 5 minutes using a personal laptop. Furthermore, our proposed approach achieves comparable or better imitation learning performance than state-of-the-art methods that only find a point estimate of the reward function. Finally, we show that our approach enables efficient high-confidence policy performance bounds. We show that these high-confidence performance bounds can be used to rank the performance and risk of a variety of evaluation policies, despite not having samples of the reward function. We also show evidence that high-confidence performance bounds can be used to detect reward hacking in complex imitation learning problems.

Local Nonparametric Meta-Learning

A central goal of meta-learning is to find a learning rule that enables fast adaptation across a set of tasks, by learning the appropriate inductive bias for that set. Most meta-learning algorithms try to find a \textit{global} learning rule that encodes this inductive bias. However, a global learning rule represented by a fixed-size representation is prone to meta-underfitting or -overfitting since the right representational power for a task set is difficult to choose a priori. Even when chosen correctly, we show that global, fixed-size representations often fail when confronted with certain types of out-of-distribution tasks, even when the same inductive bias is appropriate. To address these problems, we propose a novel nonparametric meta-learning algorithm that utilizes a meta-trained local learning rule, building on recent ideas in attention-based and functional gradient-based meta-learning. In several meta-regression problems, we show improved meta-generalization results using our local, nonparametric approach and achieve state-of-the-art results in the robotics benchmark, Omnipush.

Deep Bayesian Reward Learning from Preferences

Bayesian inverse reinforcement learning (IRL) methods are ideal for safe imitation learning, as they allow a learning agent to reason about reward uncertainty and the safety of a learned policy. However, Bayesian IRL is computationally intractable for high-dimensional problems because each sample from the posterior requires solving an entire Markov Decision Process (MDP). While there exist non-Bayesian deep IRL methods, these methods typically infer point estimates of reward functions, precluding rigorous safety and uncertainty analysis. We propose Bayesian Reward Extrapolation (B-REX), a highly efficient, preference-based Bayesian reward learning algorithm that scales to high-dimensional, visual control tasks. Our approach uses successor feature representations and preferences over demonstrations to efficiently generate samples from the posterior distribution over the demonstrator's reward function without requiring an MDP solver. Using samples from the posterior, we demonstrate how to calculate high-confidence bounds on policy performance in the imitation learning setting, in which the ground-truth reward function is unknown. We evaluate our proposed approach on the task of learning to play Atari games via imitation learning from pixel inputs, with no access to the game score. We demonstrate that B-REX learns imitation policies that are competitive with a state-of-the-art deep imitation learning method that only learns a point estimate of the reward function. Furthermore, we demonstrate that samples from the posterior generated via B-REX can be used to compute high-confidence performance bounds for a variety of evaluation policies. We show that high-confidence performance bounds are useful for accurately ranking different evaluation policies when the reward function is unknown. We also demonstrate that high-confidence performance bounds may be useful for detecting reward hacking.

Learning Hybrid Object Kinematics for Efficient Hierarchical Planning Under Uncertainty

Sudden changes in the dynamics of robotic tasks, such as contact with an object or the latching of a door, are often viewed as inconvenient discontinuities that make manipulation difficult. However, when these transitions are well-understood, they can be leveraged to reduce uncertainty or aid manipulation---for example, wiggling a screw to determine if it is fully inserted or not. Current model-free reinforcement learning approaches require large amounts of data to learn to leverage such dynamics, scale poorly as problem complexity grows, and do not transfer well to significantly different problems. By contrast, hierarchical planning-based methods scale well via plan decomposition and work well on a wide variety of problems, but often rely on precise hand-specified models and task decompositions. To combine the advantages of these opposing paradigms, we propose a new method, Act-CHAMP, which (1) learns hybrid kinematics models of objects from unsegmented data, (2) leverages actions, in addition to states, to outperform a state-of-the-art observation-only inference method, and (3) does so in a manner that is compatible with efficient, hierarchical POMDP planning. Beyond simply coping with challenging dynamics, we show that our end-to-end system leverages the learned kinematics to reduce uncertainty, plan efficiently, and use objects in novel ways not encountered during training.

Understanding Teacher Gaze Patterns for Robot Learning

Human gaze is known to be a strong indicator of underlying human intentions and goals during manipulation tasks. This work studies gaze patterns of human teachers demonstrating tasks to robots and proposes ways in which such patterns can be used to enhance robot learning. Using both kinesthetic teaching and video demonstrations, we identify novel intention-revealing gaze behaviors during teaching. These prove to be informative in a variety of problems ranging from reference frame inference to segmentation of multi-step tasks. Based on our findings, we propose two proof-of-concept algorithms which show that gaze data can enhance subtask classification for a multi-step task up to 6% and reward inference and policy learning for a single-step task up to 67%. Our findings provide a foundation for a model of natural human gaze in robot learning from demonstration settings and present open problems for utilizing human gaze to enhance robot learning.

Ranking-Based Reward Extrapolation without Rankings

The performance of imitation learning is typically upper-bounded by the performance of the demonstrator. Recent empirical results show that imitation learning via ranked demonstrations allows for better-than-demonstrator performance; however, ranked demonstrations may be difficult to obtain, and little is known theoretically about when such methods can be expected to outperform the demonstrator. To address these issues, we first contribute a sufficient condition for when better-than-demonstrator performance is possible and discuss why ranked demonstrations can contribute to better-than-demonstrator performance. Building on this theory, we then introduce Disturbance-based Reward Extrapolation (D-REX), a ranking-based imitation learning method that injects noise into a policy learned through behavioral cloning to automatically generate ranked demonstrations. By generating rankings automatically, ranking-based imitation learning can be applied in traditional imitation learning settings where only unlabeled demonstrations are available. We empirically validate our approach on standard MuJoCo and Atari benchmarks and show that D-REX can utilize automatic rankings to significantly surpass the performance of the demonstrator and outperform standard imitation learning approaches. D-REX is the first imitation learning approach to achieve significant extrapolation beyond the demonstrator's performance without additional side-information or supervision, such as rewards or human preferences.