In this paper, we investigate the semantic clustering properties of deep reinforcement learning (DRL) for video games, enriching our understanding of the internal dynamics of DRL and advancing its interpretability. In this context, semantic clustering refers to the inherent capacity of neural networks to internally group video inputs based on semantic similarity. To achieve this, we propose a novel DRL architecture that integrates a semantic clustering module featuring both feature dimensionality reduction and online clustering. This module seamlessly integrates into the DRL training pipeline, addressing instability issues observed in previous t-SNE-based analysis methods and eliminating the necessity for extensive manual annotation of semantic analysis. Through experiments, we validate the effectiveness of the proposed module and the semantic clustering properties in DRL for video games. Additionally, based on these properties, we introduce new analytical methods to help understand the hierarchical structure of policies and the semantic distribution within the feature space.

In recent years, the Robotics field has initiated several efforts toward building generalist robot policies through large-scale multi-task Behavior Cloning. However, direct deployments of these policies have led to unsatisfactory performance, where the policy struggles with unseen states and tasks. How can we break through the performance plateau of these models and elevate their capabilities to new heights? In this paper, we propose FLaRe, a large-scale Reinforcement Learning fine-tuning framework that integrates robust pre-trained representations, large-scale training, and gradient stabilization techniques. Our method aligns pre-trained policies towards task completion, achieving state-of-the-art (SoTA) performance both on previously demonstrated and on entirely novel tasks and embodiments. Specifically, on a set of long-horizon mobile manipulation tasks, FLaRe achieves an average success rate of 79.5% in unseen environments, with absolute improvements of +23.6% in simulation and +30.7% on real robots over prior SoTA methods. By utilizing only sparse rewards, our approach can enable generalizing to new capabilities beyond the pretraining data with minimal human effort. Moreover, we demonstrate rapid adaptation to new embodiments and behaviors with less than a day of fine-tuning. Videos can be found on the project website at https://robot-flare.github.io/

To obtain the optimal constraints in complex environments, Inverse Constrained Reinforcement Learning (ICRL) seeks to recover these constraints from expert demonstrations in a data-driven manner. Existing ICRL algorithms collect training samples from an interactive environment. However, the efficacy and efficiency of these sampling strategies remain unknown. To bridge this gap, we introduce a strategic exploration framework with guaranteed efficiency. Specifically, we define a feasible constraint set for ICRL problems and investigate how expert policy and environmental dynamics influence the optimality of constraints. Motivated by our findings, we propose two exploratory algorithms to achieve efficient constraint inference via 1) dynamically reducing the bounded aggregate error of cost estimation and 2) strategically constraining the exploration policy. Both algorithms are theoretically grounded with tractable sample complexity. We empirically demonstrate the performance of our algorithms under various environments. Constrained Reinforcement Learning (CRL) addresses sequential decision-making problems within safety constraints and achieves considerable success in various safety-critical applications (Gu et al., 2022). However, in many real-world environments, such as robot control (García & Shafie, 2020; Thomas et al., 2021) and autonomous driving (Krasowski et al., 2020), specifying the exact constraint that can consistently guarantee the safe control is challenging, which is further exacerbated when the ground-truth constraint is time-varying and context-dependent.

We study off-dynamics Reinforcement Learning (RL), where the policy training and deployment environments are different. To deal with this environmental perturbation, we focus on learning policies robust to uncertainties in transition dynamics under the framework of distributionally robust Markov decision processes (DRMDPs), where the nominal and perturbed dynamics are linear Markov Decision Processes. We propose a novel algorithm We-DRIVE-U that enjoys an average suboptimality $\widetilde{\mathcal{O}}\big({d H \cdot \min \{1/{\rho}, H\}/\sqrt{K} }\big)$, where $K$ is the number of episodes, $H$ is the horizon length, $d$ is the feature dimension and $\rho$ is the uncertainty level. This result improves the state-of-the-art by $\mathcal{O}(dH/\min\{1/\rho,H\})$. We also construct a novel hard instance and derive the first information-theoretic lower bound in this setting, which indicates our algorithm is near-optimal up to $\mathcal{O}(\sqrt{H})$ for any uncertainty level $\rho\in(0,1]$. Our algorithm also enjoys a 'rare-switching' design, and thus only requires $\mathcal{O}(dH\log(1+H^2K))$ policy switches and $\mathcal{O}(d^2H\log(1+H^2K))$ calls for oracle to solve dual optimization problems, which significantly improves the computational efficiency of existing algorithms for DRMDPs, whose policy switch and oracle complexities are both $\mathcal{O}(K)$.

Model-based reinforcement learning has attracted much attention due to its high sample efficiency and is expected to be applied to real-world robotic applications. In the real world, as unobservable disturbances can lead to unexpected situations, robot policies should be taken to improve not only control performance but also robustness. Adversarial learning is an effective way to improve robustness, but excessive adversary would increase the risk of malfunction, and make the control performance too conservative. Therefore, this study addresses a new adversarial learning framework to make reinforcement learning robust moderately and not conservative too much. To this end, the adversarial learning is first rederived with variational inference. In addition, light robustness, which allows for maximizing robustness within an acceptable performance degradation, is utilized as a constraint. As a result, the proposed framework, so-called LiRA, can automatically adjust adversary level, balancing robustness and conservativeness. The expected behaviors of LiRA are confirmed in numerical simulations. In addition, LiRA succeeds in learning a force-reactive gait control of a quadrupedal robot only with real-world data collected less than two hours.

A primary challenge for visual-based Reinforcement Learning (RL) is to generalize effectively across unseen environments. Although previous studies have explored different auxiliary tasks to enhance generalization, few adopt image reconstruction due to concerns about exacerbating overfitting to task-irrelevant features during training. Perceiving the pre-eminence of image reconstruction in representation learning, we propose SMG (Separated Models for Generalization), a novel approach that exploits image reconstruction for generalization. SMG introduces two model branches to extract task-relevant and task-irrelevant representations separately from visual observations via cooperatively reconstruction. Built upon this architecture, we further emphasize the importance of task-relevant features for generalization. Specifically, SMG incorporates two additional consistency losses to guide the agent's focus toward task-relevant areas across different scenarios, thereby achieving free from overfitting. Extensive experiments in DMC demonstrate the SOTA performance of SMG in generalization, particularly excelling in video-background settings. Evaluations on robotic manipulation tasks further confirm the robustness of SMG in real-world applications. Source code is available at https://anonymous.4open.science/r/SMG/.

Deep Reinforcement Learning (RL) has shown remarkable success in robotics with complex and heterogeneous dynamics. However, its vulnerability to unknown disturbances and adversarial attacks remains a significant challenge. In this paper, we propose a robust policy training framework that integrates model-based control principles with adversarial RL training to improve robustness without the need for external black-box adversaries. Our approach introduces a novel Hamilton-Jacobi reachability-guided disturbance for adversarial RL training, where we use interpretable worst-case or near-worst-case disturbances as adversaries against the robust policy. We evaluated its effectiveness across three distinct tasks: a reach-avoid game in both simulation and real-world settings, and a highly dynamic quadrotor stabilization task in simulation. We validate that our learned critic network is consistent with the ground-truth HJ value function, while the policy network shows comparable performance with other learning-based methods.

The high cost of real-world data for robotics Reinforcement Learning (RL) leads to the wide usage of simulators. Despite extensive work on building better dynamics models for simulators to match with the real world, there is another, often-overlooked mismatch between simulations and the real world, namely the distribution of available training tasks. Such a mismatch is further exacerbated by existing curriculum learning techniques, which automatically vary the simulation task distribution without considering its relevance to the real world. Considering these challenges, we posit that curriculum learning for robotics RL needs to be grounded in real-world task distributions. To this end, we propose Grounded Curriculum Learning (GCL), which aligns the simulated task distribution in the curriculum with the real world, as well as explicitly considers what tasks have been given to the robot and how the robot has performed in the past. We validate GCL using the BARN dataset on complex navigation tasks, achieving a 6.8% and 6.5% higher success rate compared to a state-of-the-art CL method and a curriculum designed by human experts, respectively. These results show that GCL can enhance learning efficiency and navigation performance by grounding the simulation task distribution in the real world within an adaptive curriculum.

We present the Duke Humanoid, an open-source 10-degrees-of-freedom humanoid, as an extensible platform for locomotion research. The design mimics human physiology, with minimized leg distances and symmetrical body alignment in the frontal plane to maintain static balance with straight knees. We develop a reinforcement learning policy that can be deployed zero-shot on the hardware for velocity-tracking walking tasks. Additionally, to enhance energy efficiency in locomotion, we propose an end-to-end reinforcement learning algorithm that encourages the robot to leverage passive dynamics. Our experiment results show that our passive policy reduces the cost of transport by up to $50\%$ in simulation and $31\%$ in real-world testing. Our website is http://generalroboticslab.com/DukeHumanoidv1/ .

In this paper, we propose an adaptive event-triggered reinforcement learning control for continuous-time nonlinear systems, subject to bounded uncertainties, characterized by complex interactions. Specifically, the proposed method is capable of jointly learning both the control policy and the communication policy, thereby reducing the number of parameters and computational overhead when learning them separately or only one of them. By augmenting the state space with accrued rewards that represent the performance over the entire trajectory, we show that accurate and efficient determination of triggering conditions is possible without the need for explicit learning triggering conditions, thereby leading to an adaptive non-stationary policy. Finally, we provide several numerical examples to demonstrate the effectiveness of the proposed approach.