This work studies the statistical limits of uniform convergence for offline policy evaluation (OPE) problems with model-based methods (for episodic MDP) and provides a unified framework towards optimal learning for several well-motivated offline tasks. Uniform OPE \sup_\Pi Q \pi-\hat{Q} \pi \epsilon is a stronger measure than the point-wise OPE and ensures offline learning when \Pi contains all policies (the global class). In this paper, we establish an \Omega(H 2 S/d_m\epsilon 2) lower bound (over model-based family) for the global uniform OPE and our main result establishes an upper bound of \tilde{O}(H 2/d_m\epsilon 2) for the \emph{local} uniform convergence that applies to all \emph{near-empirically optimal} policies for the MDPs with \emph{stationary} transition. Here d_m is the minimal marginal state-action probability. Critically, the highlight in achieving the optimal rate \tilde{O}(H 2/d_m\epsilon 2) is our design of \emph{singleton absorbing MDP}, which is a new sharp analysis tool that works with the model-based approach.

This paper introduces Honor of Kings Arena, a reinforcement learning (RL) environment based on the Honor of Kings, one of the world's most popular games at present. Compared to other environments studied in most previous work, ours presents new generalization challenges for competitive reinforcement learning. It is a multi-agent problem with one agent competing against its opponent; and it requires the generalization ability as it has diverse targets to control and diverse opponents to compete with. We describe the observation, action, and reward specifications for the Honor of Kings domain and provide an open-source Python-based interface for communicating with the game engine. We provide twenty target heroes with a variety of tasks in Honor of Kings Arena and present initial baseline results for RL-based methods with feasible computing resources. Finally, we showcase the generalization challenges imposed by Honor of Kings Arena and possible remedies to the challenges.

Mobile robots rely on maps to navigate through an environment. In the absence of any map, the robots must build the map online from partial observations as they move in the environment. Traditional methods build a map using only direct observations. In contrast, humans identify patterns in the observed environment and make informed guesses about what to expect ahead. Modeling these patterns explicitly is difficult due to the complexity of the environments. However, these complex models can be approximated well using learning-based methods in conjunction with large training data. By extracting patterns, robots can use direct observations and predictions of what lies ahead to better navigate an unknown environment. In this dissertation, we present several learning-based methods to equip mobile robots with prediction capabilities for efficient and safer operation. In the first part of the dissertation, we learn to predict using geometrical and structural patterns in the environment. Partially observed maps provide invaluable cues for accurately predicting the unobserved areas. We first demonstrate the capability of general learning-based approaches to model these patterns for a variety of overhead map modalities. Then we employ task-specific learning for faster navigation in indoor environments by predicting 2D occupancy in the nearby regions. This idea is further extended to 3D point cloud representation for object reconstruction. Predicting the shape of the full object from only partial views, our approach paves the way for efficient next-best-view planning. In the second part of the dissertation, we learn to predict using spatiotemporal patterns in the environment. We focus on dynamic tasks such as target tracking and coverage where we seek decentralized coordination between robots. We first show how graph neural networks can be used for more scalable and faster inference.

Humanoid robotics faces significant challenges in achieving stable locomotion and recovering from falls in dynamic environments. Traditional methods, such as Model Predictive Control (MPC) and Key Frame Based (KFB) routines, either require extensive fine-tuning or lack real-time adaptability. This paper introduces FRASA, a Deep Reinforcement Learning (DRL) agent that integrates fall recovery and stand up strategies into a unified framework. Leveraging the Cross-Q algorithm, FRASA significantly reduces training time and offers a versatile recovery strategy that adapts to unpredictable disturbances. Comparative tests on Sigmaban humanoid robots demonstrate FRASA superior performance against the KFB method deployed in the RoboCup 2023 by the Rhoban Team, world champion of the KidSize League.

Building deep reinforcement learning (RL) agents that find a good policy with few samples has proven notoriously challenging. To achieve sample efficiency, recent work has explored updating neural networks with large numbers of gradient steps for every new sample. While such high update-to-data (UTD) ratios have shown strong empirical performance, they also introduce instability to the training process. Previous approaches need to rely on periodic neural network parameter resets to address this instability, but restarting the training process is infeasible in many real-world applications and requires tuning the resetting interval. In this paper, we focus on one of the core difficulties of stable training with limited samples: the inability of learned value functions to generalize to unobserved on-policy actions. We mitigate this issue directly by augmenting the off-policy RL training process with a small amount of data generated from a learned world model. Our method, Model-Augmented Data for Temporal Difference learning (MAD-TD) uses small amounts of generated data to stabilize high UTD training and achieve competitive performance on the most challenging tasks in the DeepMind control suite. Our experiments further highlight the importance of employing a good model to generate data, MAD-TD's ability to combat value overestimation, and its practical stability gains for continued learning.

The manufacturing sector like many other sectors was recently affected by workforce shortages, a problem that automation and robotics can heavily minimize Kugler (2022). Simultaneously, Reinforcement learning (RL) offers a promising solution where robots can learn to perform tasks through interaction and feedback from the environment Singh et al. (2022). However, despite their success in numerous simulation environments, we still don't see many real-world deployments of RL robotic solutions. In fact, many researchers either oversimplify the targeted real-world scenario such as Wu et al. (2023) or do not even evaluate their model in physical robots Lu et al. (2022); Na et al. (2022). It is known that training RL policies directly in real robots can be expensive, timeconsuming, labor-intensive, and maybe even dangerous, that is why it makes sense to try to leverage training in simulation.

Learning from Demonstrations, particularly from biological experts like humans and animals, often encounters significant data acquisition challenges. While recent approaches leverage internet videos for learning, they require complex, task-specific pipelines to extract and retarget motion data for the agent. In this work, we introduce a language-model-assisted bi-level programming framework that enables a reinforcement learning agent to directly learn its reward from internet videos, bypassing dedicated data preparation. The framework includes two levels: an upper level where a vision-language model (VLM) provides feedback by comparing the learner's behavior with expert videos, and a lower level where a large language model (LLM) translates this feedback into reward updates. The VLM and LLM collaborate within this bi-level framework, using a "chain rule" approach to derive a valid search direction for reward learning. We validate the method for reward learning from YouTube videos, and the results have shown that the proposed method enables efficient reward design from expert videos of biological agents for complex behavior synthesis.

In multi-agent reinforcement learning (MARL), parameter sharing is commonly employed to enhance sample efficiency. However, the popular approach of full parameter sharing often leads to homogeneous policies among agents, potentially limiting the performance benefits that could be derived from policy diversity. To address this critical limitation, we introduce \emph{Kaleidoscope}, a novel adaptive partial parameter sharing scheme that fosters policy heterogeneity while still maintaining high sample efficiency. Specifically, Kaleidoscope maintains one set of common parameters alongside multiple sets of distinct, learnable masks for different agents, dictating the sharing of parameters. It promotes diversity among policy networks by encouraging discrepancy among these masks, without sacrificing the efficiencies of parameter sharing. This design allows Kaleidoscope to dynamically balance high sample efficiency with a broad policy representational capacity, effectively bridging the gap between full parameter sharing and non-parameter sharing across various environments. We further extend Kaleidoscope to critic ensembles in the context of actor-critic algorithms, which could help improve value estimations.Our empirical evaluations across extensive environments, including multi-agent particle environment, multi-agent MuJoCo and StarCraft multi-agent challenge v2, demonstrate the superior performance of Kaleidoscope compared with existing parameter sharing approaches, showcasing its potential for performance enhancement in MARL. The code is publicly available at \url{https://github.com/LXXXXR/Kaleidoscope}.

Learning a latent dynamics model provides a task-agnostic representation of an agent's understanding of its environment. Leveraging this knowledge for model-based reinforcement learning holds the potential to improve sample efficiency over model-free methods by learning inside imagined rollouts. Furthermore, because the latent space serves as input to behavior models, the informative representations learned by the world model facilitate efficient learning of desired skills. Most existing methods rely on holistic representations of the environment's state. In contrast, humans reason about objects and their interactions, forecasting how actions will affect specific parts of their surroundings. Inspired by this, we propose Slot-Attention for Object-centric Latent Dynamics (SOLD), a novel algorithm that learns object-centric dynamics models in an unsupervised manner from pixel inputs. We demonstrate that the structured latent space not only improves model interpretability but also provides a valuable input space for behavior models to reason over. Our results show that SOLD outperforms DreamerV3, a state-of-the-art model-based RL algorithm, across a range of benchmark robotic environments that evaluate for both relational reasoning and low-level manipulation capabilities. Videos are available at https://slot-latent-dynamics.github.io/.

High-resolution (HR) contact surface information is essential for robotic grasping and precise manipulation tasks. However, it remains a challenge for current taxel-based sensors to obtain HR tactile information. In this paper, we focus on utilizing low-resolution (LR) tactile sensors to reconstruct the localized, dense, and HR representation of contact surfaces. In particular, we build a Gaussian triaxial tactile sensor degradation model and propose a tactile pattern reconstruction framework based on the Kalman filter. This framework enables the reconstruction of 2-D HR contact surface shapes using collected LR tactile sequences. In addition, we present an active exploration strategy to enhance the reconstruction efficiency. We evaluate the proposed method in real-world scenarios with comparison to existing prior-information-based approaches. Experimental results confirm the efficiency of the proposed approach and demonstrate satisfactory reconstructions of complex contact surface shapes. Code: https://github.com/wmtlab/tactileAR