Multi-Fidelity Policy Gradient Algorithms

Many reinforcement learning (RL) algorithms require large amounts of data, prohibiting their use in applications where frequent interactions with operational systems are infeasible, or high-fidelity simulations are expensive or unavailable. Meanwhile, low-fidelity simulators--such as reduced-order models, heuristic reward functions, or generative world models--can cheaply provide useful data for RL training, even if they are too coarse for direct sim-to-real transfer. We propose multi-fidelity policy gradients (MFPGs), an RL framework that mixes a small amount of data from the target environment with a large volume of low-fidelity simulation data to form unbiased, reduced-variance estimators (control variates) for on-policy policy gradients. We instantiate the framework by developing multi-fidelity variants of two policy gradient algorithms: REINFORCE and proximal policy optimization. Experimental results across a suite of simulated robotics benchmark problems demonstrate that when target-environment samples are limited, MFPG achieves up to 3.9x higher reward and improves training stability when compared to baselines that only use high-fidelity data. Moreover, even when the baselines are given more high-fidelity samples--up to 10x as many interactions with the target environment--MFPG continues to match or outperform them. Finally, we observe that MFPG is capable of training effective policies even when the low-fidelity environment is drastically different from the target environment. MFPG thus not only offers a novel paradigm for efficient sim-to-real transfer but also provides a principled approach to managing the trade-off between policy performance and data collection costs.

Stealing That Free Lunch: Exposing the Limits of Dyna-Style Reinforcement Learning

Dyna-style off-policy model-based reinforcement learning (DMBRL) algorithms are a family of techniques for generating synthetic state transition data and thereby enhancing the sample efficiency of off-policy RL algorithms. This paper identifies and investigates a surprising performance gap observed when applying DMBRL algorithms across different benchmark environments with proprioceptive observations. We show that, while DMBRL algorithms perform well in OpenAI Gym, their performance can drop significantly in DeepMind Control Suite (DMC), even though these settings offer similar tasks and identical physics backends. Modern techniques designed to address several key issues that arise in these settings do not provide a consistent improvement across all environments, and overall our results show that adding synthetic rollouts to the training process -- the backbone of Dyna-style algorithms -- significantly degrades performance across most DMC environments. Our findings contribute to a deeper understanding of several fundamental challenges in model-based RL and show that, like many optimization fields, there is no free lunch when evaluating performance across diverse benchmarks in RL.

Inferring Short-Sightedness in Dynamic Noncooperative Games

Dynamic game theory is an increasingly popular tool for modeling multi-agent, e.g. human-robot, interactions. Game-theoretic models presume that each agent wishes to minimize a private cost function that depends on others' actions. These games typically evolve over a fixed time horizon, which specifies the degree to which all agents care about the distant future. In practical settings, however, decision-makers may vary in their degree of short-sightedness. We conjecture that quantifying and estimating each agent's short-sightedness from online data will enable safer and more efficient interactions with other agents. To this end, we frame this inference problem as an inverse dynamic game. We consider a specific parametrization of each agent's objective function that smoothly interpolates myopic and farsighted planning. Games of this form are readily transformed into parametric mixed complementarity problems; we exploit the directional differentiability of solutions to these problems with respect to their hidden parameters in order to solve for agents' short-sightedness. We conduct several experiments simulating human behavior at a real-world crosswalk. The results of these experiments clearly demonstrate that by explicitly inferring agents' short-sightedness, we can recover more accurate game-theoretic models, which ultimately allow us to make better predictions of agents' behavior. Specifically, our results show up to a 30% more accurate prediction of myopic behavior compared to the baseline.

Dense Dynamics-Aware Reward Synthesis: Integrating Prior Experience with Demonstrations

Many continuous control problems can be formulated as sparse-reward reinforcement learning (RL) tasks. In principle, online RL methods can automatically explore the state space to solve each new task. However, discovering sequences of actions that lead to a non-zero reward becomes exponentially more difficult as the task horizon increases. Manually shaping rewards can accelerate learning for a fixed task, but it is an arduous process that must be repeated for each new environment. We introduce a systematic reward-shaping framework that distills the information contained in 1) a task-agnostic prior data set and 2) a small number of task-specific expert demonstrations, and then uses these priors to synthesize dense dynamics-aware rewards for the given task. This supervision substantially accelerates learning in our experiments, and we provide analysis demonstrating how the approach can effectively guide online learning agents to faraway goals.

Policies with Sparse Inter-Agent Dependencies in Dynamic Games: A Dynamic Programming Approach

Common feedback strategies in multi-agent dynamic games require all players' state information to compute control strategies. However, in real-world scenarios, sensing and communication limitations between agents make full state feedback expensive or impractical, and such strategies can become fragile when state information from other agents is inaccurate. To this end, we propose a regularized dynamic programming approach for finding sparse feedback policies that selectively depend on the states of a subset of agents in dynamic games. The proposed approach solves convex adaptive group Lasso problems to compute sparse policies approximating Nash equilibrium solutions. We prove the regularized solutions' asymptotic convergence to a neighborhood of Nash equilibrium policies in linear-quadratic (LQ) games. We extend the proposed approach to general non-LQ games via an iterative algorithm. Empirical results in multi-robot interaction scenarios show that the proposed approach effectively computes feedback policies with varying sparsity levels. When agents have noisy observations of other agents' states, simulation results indicate that the proposed regularized policies consistently achieve lower costs than standard Nash equilibrium policies by up to 77% for all interacting agents whose costs are coupled with other agents' states.

Learning to Walk from Three Minutes of Real-World Data with Semi-structured Dynamics Models

Traditionally, model-based reinforcement learning (MBRL) methods exploit neural networks as flexible function approximators to represent a priori unknown environment dynamics. However, training data are typically scarce in practice, and these black-box models often fail to generalize. Modeling architectures that leverage known physics can substantially reduce the complexity of system-identification, but break down in the face of complex phenomena such as contact. We introduce a novel framework for learning semi-structured dynamics models for contact-rich systems which seamlessly integrates structured first principles modeling techniques with black-box auto-regressive models. Specifically, we develop an ensemble of probabilistic models to estimate external forces, conditioned on historical observations and actions, and integrate these predictions using known Lagrangian dynamics. With this semi-structured approach, we can make accurate long-horizon predictions with substantially less data than prior methods. We leverage this capability and propose Semi-Structured Reinforcement Learning (SSRL) a simple model-based learning framework which pushes the sample complexity boundary for real-world learning. We validate our approach on a real-world Unitree Go1 quadruped robot, learning dynamic gaits -- from scratch -- on both hard and soft surfaces with just a few minutes of real-world data. Video and code are available at: https://sites.google.com/utexas.edu/ssrl

Learning responsibility allocations for multi-agent interactions: A differentiable optimization approach with control barrier functions

From autonomous driving to package delivery, ensuring safe yet efficient multi-agent interaction is challenging as the interaction dynamics are influenced by hard-to-model factors such as social norms and contextual cues. Understanding these influences can aid in the design and evaluation of socially-aware autonomous agents whose behaviors are aligned with human values. In this work, we seek to codify factors governing safe multi-agent interactions via the lens of responsibility, i.e., an agent's willingness to deviate from their desired control to accommodate safe interaction with others. Specifically, we propose a data-driven modeling approach based on control barrier functions and differentiable optimization that efficiently learns agents' responsibility allocation from data. We demonstrate on synthetic and real-world datasets that we can obtain an interpretable and quantitative understanding of how much agents adjust their behavior to ensure the safety of others given their current environment.

Decomposing Control Lyapunov Functions for Efficient Reinforcement Learning

Recent methods using Reinforcement Learning (RL) have proven to be successful for training intelligent agents in unknown environments. However, RL has not been applied widely in real-world robotics scenarios. This is because current state-of-the-art RL methods require large amounts of data to learn a specific task, leading to unreasonable costs when deploying the agent to collect data in real-world applications. In this paper, we build from existing work that reshapes the reward function in RL by introducing a Control Lyapunov Function (CLF), which is demonstrated to reduce the sample complexity. Still, this formulation requires knowing a CLF of the system, but due to the lack of a general method, it is often a challenge to identify a suitable CLF. Existing work can compute low-dimensional CLFs via a Hamilton-Jacobi reachability procedure. However, this class of methods becomes intractable on high-dimensional systems, a problem that we address by using a system decomposition technique to compute what we call Decomposed Control Lyapunov Functions (DCLFs). We use the computed DCLF for reward shaping, which we show improves RL performance. Through multiple examples, we demonstrate the effectiveness of this approach, where our method finds a policy to successfully land a quadcopter in less than half the amount of real-world data required by the state-of-the-art Soft-Actor Critic algorithm.

Auto-Encoding Bayesian Inverse Games

When multiple agents interact in a common environment, each agent's actions impact others' future decisions, and noncooperative dynamic games naturally capture this coupling. In interactive motion planning, however, agents typically do not have access to a complete model of the game, e.g., due to unknown objectives of other players. Therefore, we consider the inverse game problem, in which some properties of the game are unknown a priori and must be inferred from observations. Existing maximum likelihood estimation (MLE) approaches to solve inverse games provide only point estimates of unknown parameters without quantifying uncertainty, and perform poorly when many parameter values explain the observed behavior. To address these limitations, we take a Bayesian perspective and construct posterior distributions of game parameters. To render inference tractable, we employ a variational autoencoder (VAE) with an embedded differentiable game solver. This structured VAE can be trained from an unlabeled dataset of observed interactions, naturally handles continuous, multi-modal distributions, and supports efficient sampling from the inferred posteriors without computing game solutions at runtime. Extensive evaluations in simulated driving scenarios demonstrate that the proposed approach successfully learns the prior and posterior objective distributions, provides more accurate objective estimates than MLE baselines, and facilitates safer and more efficient game-theoretic motion planning.

An Investigation of Time Reversal Symmetry in Reinforcement Learning

One of the fundamental challenges associated with reinforcement learning (RL) is that collecting sufficient data can be both time-consuming and expensive. In this paper, we formalize a concept of time reversal symmetry in a Markov decision process (MDP), which builds upon the established structure of dynamically reversible Markov chains (DRMCs) and time-reversibility in classical physics. Specifically, we investigate the utility of this concept in reducing the sample complexity of reinforcement learning. We observe that utilizing the structure of time reversal in an MDP allows every environment transition experienced by an agent to be transformed into a feasible reverse-time transition, effectively doubling the number of experiences in the environment. To test the usefulness of this newly synthesized data, we develop a novel approach called time symmetric data augmentation (TSDA) and investigate its application in both proprioceptive and pixel-based state within the realm of off-policy, model-free RL. Empirical evaluations showcase how these synthetic transitions can enhance the sample efficiency of RL agents in time reversible scenarios without friction or contact. We also test this method in more realistic environments where these assumptions are not globally satisfied. We find that TSDA can significantly degrade sample efficiency and policy performance, but can also improve sample efficiency under the right conditions. Ultimately we conclude that time symmetry shows promise in enhancing the sample efficiency of reinforcement learning and provide guidance when the environment and reward structures are of an appropriate form for TSDA to be employed effectively.