The paper proposes a framework to reason about object dynamics for faster reinforcement learning.
problem Current reinforcement learning approaches lack prior knowledge about the environment, limiting efficiency.
method Integrates object dynamics and behavior into reinforcement learning to improve efficiency.
result Demonstrates the need for reasoning about object behavior and dynamics, leading to faster learning.
Novel PO algorithms improve LLM alignment tasks.
problem Evaluating PO algorithms on LLM alignment is costly and noisy.
method Designed a diagnostic suite of MuJoCo tasks and datasets, used evolutionary strategies to discover specialized PO algorithms.
result Proposed MPO algorithms significantly outperform existing PO algorithms in LLM alignment.
Many mobile robots rely on 2D laser scanners for localization, mapping, and navigation. However, those sensors are unable to correctly provide distance to obstacles such as glass panels and tables whose actual occupancy is invisible at the height the sensor is measuring. In this work, instead of estimating the distance…
MVPI framework optimizes risk in reinforcement learning, improving performance in robot simulations.
problem Optimizing risk in reinforcement learning control problems.
method Mean-Variance Policy Iteration (MVPI) framework for risk-averse control in MDPs.
result Risk-averse TD3 outperforms previous methods in robot simulation tasks.
PRISM provides real-time SLAM with uncertainty estimates for agent and map states.
problem Lack of uncertainty estimates and real-time capability in SLAM.
method Combines differentiable rendering and 6-DoF dynamics, uses approximations for Bayesian inference.
result Runs at 10Hz real-time with similar accuracy to state-of-the-art SLAM.
New loss function equivalence reveals PER's uniform sampling can be improved.
problem Improving Prioritized Experience Replay (PER) for better learning efficiency.
method Transforming non-uniformly sampled data loss functions into uniformly sampled ones.
result Some environments can replace PER with a new loss function without performance loss.
Meta-gradient D4PG optimizes performance and constraint adherence in RL.
problem Balancing performance and adherence to complex constraints in RL.
method Uses meta-gradients to find a balance between expected return and minimizing constraint violations.
result Meta-gradient D4PG consistently outperforms baselines across MuJoCo domains.
Deep RL approach improves MIS for complex environments.
problem Improving off-policy evaluation for complex environments.
method Uses successor representation from deep RL to decouple reward and dynamics.
result Empirically stable and applicable to high-dimensional domains.
CVRL tackles complex visual observations in reinforcement learning.
problem Complex visual observations in natural environments.
method Contrastive Variational Reinforcement Learning (CVRL) learns a contrastive variational model by maximizing mutual information between latent states and observations.
result CVRL achieves comparable performance with state-of-the-art model-based DRL methods and significantly outperforms them on tasks with complex observations.
Classifier-based AI safety gates fail in self-improvement, even with advanced verification methods.
problem Maintaining reliable oversight of AI systems as they improve over iterations.
method Comprehensive empirical testing on neural controllers and MuJoCo benchmarks, using various classifiers and verification methods.
result Classifier-based safety gates fail in maintaining reliable oversight, even with advanced verification methods.
In this paper, we present a method for learning from video demonstrations by using human feedback to construct a mapping between the standard representation of the agent and the visual representation of the demonstration. In this way, we leverage the advantages of both these representations, i.e., we learn the policy u…
Simplifies RL training with fewer techniques, reducing bias and instability.
problem Training instabilities and high sample complexity in RL.
method Introduced a simple deterministic policy gradient, used propensity estimation, and delayed policy updates.
result Improved performance and reduced sample complexity through these techniques.
DiCE uses diverse agents to explore and learn, avoiding local minima.
problem Local minima in RL due to limited exploration and correlated behavior.
method DiCE employs a group of heterogeneous agents to explore simultaneously and share experiences, with a diversity regularization mechanism.
result DiCE achieves substantial improvement over baselines in MuJoCo locomotion tasks.
PWIL learns agent behavior from expert using Wasserstein distance.
problem Learning agent behavior from expert demonstrations efficiently.
method Primal Wasserstein Imitation Learning (PWIL) with offline reward function.
result PWIL achieves expert behavior on MuJoCo tasks with minimal interactions.
Dual-structured method improves cross-domain imitation learning.
problem Learning policies in a target domain from source domain expert demonstrations.
method Extracts domain and policy expertness features, synthesizes new features.
result Superior performance compared to other algorithms in MuJoCo tasks.
MAGE optimizes policies using action gradients from model-based learning.
problem Lack of direct gradient information from critics in actor-critic methods.
method Model-based actor-critic algorithm that learns action-value gradient.
result MAGE outperforms model-free and model-based baselines on continuous control tasks.
ESPD improves learning efficiency in sparse reward reinforcement learning.
problem Sparse reward reinforcement learning challenges.
method Evolutionary Stochastic Policy Distillation (ESPD) based on drifted random walk insight.
result High learning efficiency demonstrated in MuJoCo robotics control suite experiments.
New approach relaxes inductive biases of physics-inspired NNs for better performance.
problem Challenges in applying physics-inspired NNs to real-world systems.
method Examined and relaxed inductive biases of Hamiltonian NNs, improving performance on non-conservative systems.
result Improved performance on practical, non-conservative systems by relaxing inductive biases.
Algorithm learns actions from past states in complex tasks.
problem Learning policies from human feedback is expensive.
method Combining learned feature encoder with inverse models to simulate past actions.
result Algorithm can infer specific skills from single state.
Paper introduces DTAE to optimize RL algorithms, balancing exploration and exploitation.
problem Balancing exploration and exploitation in reinforcement learning.
method Soft policy optimization with entropy and dual-track advantage estimator (DTAE).
result DTAE accelerates RL algorithm convergence and improves performance.
We present the OpenAI Remote Rendering Backend (ORRB), a system that allows fast and customizable rendering of robotics environments. It is based on the Unity3d game engine and interfaces with the MuJoCo physics simulation library. ORRB was designed with visual domain randomization in mind. It is optimized for cloud de…
New methods improve deep reinforcement learning by accelerating credit assignment.
problem Challenges in achieving fast and stable off-policy learning in deep reinforcement learning.
method Extends the generalized PBE objective to support multistep credit assignment and derives three gradient-based methods.
result Proposed methods outperform PPO and StreamQ in MuJoCo and MinAtar environments.
AdaRL improves robust RL by adaptively adjusting policy complexity.
problem Handling epistemic uncertainty in environment dynamics.
method Bi-level optimization framework with adaptive rank adjustment.
result AdaRL outperforms existing methods on MuJoCo benchmarks.
A new method for ILO with transition model disparity using an intermediary policy.
problem Learning tasks from expert observations with different transition dynamics.
method Training an intermediary policy to match the state transitions of the expert dataset.
result Our method outperforms existing ILO approaches with transition model mismatch.
This study compares 6 imitation learning algorithms using a common dataset and hyperparameter budget.
problem Difficulty in comparing different imitation learning algorithms due to varying datasets, base RL algorithms, and evaluation settings.
method Reimplemented and updated 6 different imitation learning algorithms, using a common off-policy algorithm (SAC) and a widely-used dataset (D4RL). Evaluated on a range of expert trajectories.
result GAIL consistently performs well across different sample sizes, while AdRIL performs well with one important hyperparameter to tune and behavioral cloning remains a strong baseline when data is plentiful.
Rewards are sparse in the real world and most of today's reinforcement learning algorithms struggle with such sparsity. One solution to this problem is to allow the agent to create rewards for itself - thus making rewards dense and more suitable for learning. In particular, inspired by curious behaviour in animals, obs…
This paper proposes Self-Imitation Learning (SIL), a simple off-policy actor-critic algorithm that learns to reproduce the agent's past good decisions. This algorithm is designed to verify our hypothesis that exploiting past good experiences can indirectly drive deep exploration. Our empirical results show that SIL sig…
Soft Actor-Critic (SAC) is an off-policy actor-critic deep reinforcement learning (DRL) algorithm based on maximum entropy reinforcement learning. By combining off-policy updates with an actor-critic formulation, SAC achieves state-of-the-art performance on a range of continuous-action benchmark tasks, outperforming pr…
PPO's gradients are heavy-tailed, affecting learning; a robust estimator improves performance.
problem Heavy-tailedness of PPO gradients causing learning issues.
method Characterized heavy-tailed gradients, identified likelihood ratios and advantages as sources, proposed GMOM as a robust estimator.
result GMOM improves PPO performance without clipping tricks.
Paper tackles covariate shift in offline IL using less proficient behavior data.
problem Mitigating covariate shift in Imitation Learning with limited data coverage.
method Model-based IL from Offline Data (MILO) framework.
result MILO can combat covariate shift even with sub-optimal behavior policy data.
New state-only IL algorithm tackles MDP transition mismatch.
problem Transition dynamics mismatch between expert and imitator MDPs.
method Adversarial state-only IL with two subproblems solved iteratively.
result Effective performance improvement in transition dynamics mismatch scenarios.
Sparse rewards with different magnitudes speed up learning in model-based reinforcement learning.
problem Speeding up learning in model-based reinforcement learning methods.
method Add uniformly sampled reward values to sparse binary rewards during training.
result Training can be more efficient and faster with less variability.
Novel algorithm reduces computational burden in IRL with finite-time guarantees.
problem Efficiently recover reward function and optimal policy from expert behavior.
method Single-loop algorithm that maximizes likelihood after each policy improvement step.
result Algorithm provably converges to a stationary solution with finite-time guarantees.
This paper explores a simple regularizer for reinforcement learning by proposing Generative Adversarial Self-Imitation Learning (GASIL), which encourages the agent to imitate past good trajectories via generative adversarial imitation learning framework. Instead of directly maximizing rewards, GASIL focuses on reproduc…
FORK improves model-free reinforcement learning performance.
problem Improving model-free reinforcement learning performance.
method Introducing a new forward-looking Actor (FORK) for Actor-Critic algorithms.
result FORK significantly improves performance in various environments.
New RL method improves robustness across various conditions.
problem Training robust RL agents for diverse environments.
method Adversarial training with Langevin Dynamics.
result Consistently outperforms existing RL algorithms in generalization.
RL agents learn from a few tasks to generalize to new ones.
problem Creating efficient RL agents that can solve multiple tasks.
method GHP-MDPs model with latent variables for hidden parameters.
result State-of-the-art performance and sample-efficiency on new tasks.
Many continuous control tasks have easily formulated objectives, yet using them directly as a reward in reinforcement learning (RL) leads to suboptimal policies. Therefore, many classical control tasks guide RL training using complex rewards, which require tedious hand-tuning. We automate the reward search with AutoRL,…
Policy optimization on high-dimensional continuous control tasks exhibits its difficulty caused by the large variance of the policy gradient estimators. We present the action subspace dependent gradient (ASDG) estimator which incorporates the Rao-Blackwell theorem (RB) and Control Variates (CV) into a unified framework…
Recent advances in policy gradient methods and deep learning have demonstrated their applicability for complex reinforcement learning problems. However, the variance of the performance gradient estimates obtained from the simulation is often excessive, leading to poor sample efficiency. In this paper, we apply the stoc…
There are two halves to RL systems: experience collection time and policy learning time. For a large number of samples in rollouts, experience collection time is the major bottleneck. Thus, it is necessary to speed up the rollout generation time with multi-process architecture support. Our work, dubbed WALL-E, utilizes…
In hierarchical reinforcement learning a major challenge is determining appropriate low-level policies. We propose an unsupervised learning scheme, based on asymmetric self-play from Sukhbaatar et al. (2018), that automatically learns a good representation of sub-goals in the environment and a low-level policy that can…
In this technical report, we consider an approach that combines the PPO objective and K-FAC natural gradient optimization, for which we call PPOKFAC. We perform a range of empirical analysis on various aspects of the algorithm, such as sample complexity, training speed, and sensitivity to batch size and training epochs…
We propose a new objective, the counterfactual objective, unifying existing objectives for off-policy policy gradient algorithms in the continuing reinforcement learning (RL) setting. Compared to the commonly used excursion objective, which can be misleading about the performance of the target policy when deployed, our…
Genetic algorithms have been widely used in many practical optimization problems. Inspired by natural selection, operators, including mutation, crossover and selection, provide effective heuristics for search and black-box optimization. However, they have not been shown useful for deep reinforcement learning, possibly …
New RL approach speeds up training across tasks.
problem Training reinforcement learning models quickly and efficiently.
method Combines planning quasi-metric and task-specific aimers.
result Achieves multiple-fold speed-up on bit-flip and robotic arm tasks.
CoNES optimizes blackbox functions using convex optimization and information geometry.
problem Optimizing high-dimensional blackbox functions efficiently.
method Formulated as a convex program that adapts evolutionary strategies gradient estimates.
result Vastly outperforms conventional blackbox optimization methods on benchmarks and MuJoCo tasks.
Bayesian approach uses Gaussian process for reinforcement learning.
problem Robotic locomotion environments
method Bayesian actor-critic, model-free reinforcement learning with Gaussian process for exploration and policy optimization.
result Gaussian process method outperforms current algorithms in robotic locomotion environments.