Abstract--Towards the grand challenge of achieving humanlevel manipulation in robots, playing piano is a compelling testbed that requires strategic, precise, and flowing movements. Over the years, several works demonstrated hand-designed controllers on real world piano playing, while other works evaluated robot learning approaches on simulated piano scenarios. In this paper, we develop the first piano playing robotic system that makes use of learning approaches while also being deployed on a real world dexterous robot. Specifically, we make use of Sim2Real to train a policy in simulation using reinforcement learning before deploying the learned policy on a real world dexterous robot. In our experiments, we thoroughly evaluate the interplay between domain randomization and the accuracy of the dynamics model used in simulation. Moreover, we evaluate the robot's performance across multiple songs with varying complexity to study the generalization of our learned policy. Experimental results show that the robot can learn Playing the piano requires humans to master contact-rich to play several simple pieces successfully, after training exclusively hand movements dictated by the timing and tone they intend in simulation. This mastery is not learned quickly but through extensive practice, which requires humans to control their actions based on the haptic and auditory feedback received the natural movements of human hands. This makes it an ideal with each key pressed on the piano. In addition, human hands scenario for exploring Sim2Real transfer, where the objective are an extraordinary research subject due to their unmatched is to train an agent in simulation capable of performing in the dexterity, precision, and adaptability.

Social robot navigation is an evolving research field that aims to find efficient strategies to safely navigate dynamic environments populated by humans. A critical challenge in this domain is the accurate modeling of human motion, which directly impacts the design and evaluation of navigation algorithms. This paper presents a comparative study of two popular categories of human motion models used in social robot navigation, namely velocity-based models and force-based models. A system-theoretic representation of both model types is presented, which highlights their common feedback structure, although with different state variables. Several navigation policies based on reinforcement learning are trained and tested in various simulated environments involving pedestrian crowds modeled with these approaches. A comparative study is conducted to assess performance across multiple factors, including human motion model, navigation policy, scenario complexity and crowd density. The results highlight advantages and challenges of different approaches to modeling human behavior, as well as their role during training and testing of learning-based navigation policies. The findings offer valuable insights and guidelines for selecting appropriate human motion models when designing socially-aware robot navigation systems.

This paper proposes a design scheme of reward function that constantly evaluates both driving states and actions for applying reinforcement learning to automated driving. In the field of reinforcement learning, reward functions often evaluate whether the goal is achieved by assigning values such as +1 for success and -1 for failure. This type of reward function can potentially obtain a policy that achieves the goal, but the process by which the goal is reached is not evaluated. However, process to reach a destination is important for automated driving, such as keeping velocity, avoiding risk, retaining distance from other cars, keeping comfortable for passengers. Therefore, the reward function designed by the proposed scheme is suited for automated driving by evaluating driving process. The effects of the proposed scheme are demonstrated on simulated circuit driving and highway cruising. Asynchronous Advantage Actor-Critic is used, and models are trained under some situations for generalization. The result shows that appropriate driving positions are obtained, such as traveling on the inside of corners, and rapid deceleration to turn along sharp curves. In highway cruising, the ego vehicle becomes able to change lane in an environment where there are other vehicles with suitable deceleration to avoid catching up to a front vehicle, and acceleration so that a rear vehicle does not catch up to the ego vehicle.

Recently, reinforcement learning (RL) has been shown to greatly enhance the reasoning capabilities of large language models (LLMs), and RL-based approaches have been progressively applied to visual multimodal tasks. However, the audio modality has largely been overlooked in these developments. Thus, we conduct a series of RL explorations in audio understanding and reasoning, specifically focusing on the audio question answering (AQA) task. We leverage the group relative policy optimization (GRPO) algorithm to Qwen2-Audio-7B-Instruct, and our experiments demonstrated state-of-the-art performance on the MMAU Test-mini benchmark, achieving an accuracy rate of 64.5%. The main findings in this technical report are as follows: 1) The GRPO algorithm can be effectively applied to large audio language models (LALMs), even when the model has only 8.2B parameters; 2) With only 38k post-training samples, RL significantly outperforms supervised fine-tuning (SFT), indicating that RL-based approaches can be effective without large datasets; 3) The explicit reasoning process has not shown significant benefits for AQA tasks, and how to efficiently utilize deep thinking remains an open question for further research; 4) LALMs still lag far behind humans auditory-language reasoning, suggesting that the RL-based approaches warrant further explorations.

We study the problem of imitating an expert demonstrator in a discrete-time, continuous state-and-action control system. We show that, even if the dynamics are stable (i.e. contracting exponentially quickly), and the expert is smooth and deterministic, any smooth, deterministic imitator policy necessarily suffers error on execution that is exponentially larger, as a function of problem horizon, than the error under the distribution of expert training data. Our negative result applies to both behavior cloning and offline-RL algorithms, unless they produce highly "improper" imitator policies--those which are non-smooth, non-Markovian, or which exhibit highly state-dependent stochasticity--or unless the expert trajectory distribution is sufficiently "spread." We provide experimental evidence of the benefits of these more complex policy parameterizations, explicating the benefits of today's popular policy parameterizations in robot learning (e.g. action-chunking and Diffusion Policies). We also establish a host of complementary negative and positive results for imitation in control systems.

Mobile robot navigation in dynamic environments with pedestrian traffic is a key challenge in the development of autonomous mobile service robots. Recently, deep reinforcement learning-based methods have been actively studied and have outperformed traditional rule-based approaches owing to their optimization capabilities. Among these, methods that assume a continuous action space typically rely on a Gaussian distribution assumption, which limits the flexibility of generated actions. Meanwhile, the application of diffusion models to reinforcement learning has advanced, allowing for more flexible action distributions compared with Gaussian distribution-based approaches. In this study, we applied a diffusion-based reinforcement learning approach to social navigation and validated its effectiveness. Furthermore, by leveraging the characteristics of diffusion models, we propose an extension that enables post-training action smoothing and adaptation to static obstacle scenarios not considered during the training steps.

Scaling up self-supervised learning has driven breakthroughs in language and vision, yet comparable progress has remained elusive in reinforcement learning (RL). In this paper, we study building blocks for self-supervised RL that unlock substantial improvements in scalability, with network depth serving as a critical factor. Whereas most RL papers in recent years have relied on shallow architectures (around 2 - 5 layers), we demonstrate that increasing the depth up to 1024 layers can significantly boost performance. Our experiments are conducted in an unsupervised goal-conditioned setting, where no demonstrations or rewards are provided, so an agent must explore (from scratch) and learn how to maximize the likelihood of reaching commanded goals. Evaluated on simulated locomotion and manipulation tasks, our approach increases performance by $2\times$ - $50\times$. Increasing the model depth not only increases success rates but also qualitatively changes the behaviors learned.

Decision-making in complex, continuous multi-task environments is often hindered by the difficulty of obtaining accurate models for planning and the inefficiency of learning purely from trial and error. While precise environment dynamics may be hard to specify, human experts can often provide high-fidelity abstractions that capture the essential high-level structure of a task and user preferences in the target environment. Existing hierarchical approaches often target discrete settings and do not generalize across tasks. We propose a hierarchical reinforcement learning approach that addresses these limitations by dynamically planning over the expert-specified abstraction to generate subgoals to learn a goal-conditioned policy. To overcome the challenges of learning under sparse rewards, we shape the reward based on the optimal state value in the abstract model. This structured decision-making process enhances sample efficiency and facilitates zero-shot generalization. Our empirical evaluation on a suite of procedurally generated continuous control environments demonstrates that our approach outperforms existing hierarchical reinforcement learning methods in terms of sample efficiency, task completion rate, scalability to complex tasks, and generalization to novel scenarios.

The sim-to-real gap remains a critical challenge in robotics, hindering the deployment of algorithms trained in simulation to real-world systems. This paper introduces a novel Real-Sim-Real (RSR) loop framework leveraging differentiable simulation to address this gap by iteratively refining simulation parameters, aligning them with real-world conditions, and enabling robust and efficient policy transfer. A key contribution of our work is the design of an informative cost function that encourages the collection of diverse and representative real-world data, minimizing bias and maximizing the utility of each data point for simulation refinement. This cost function integrates seamlessly into existing reinforcement learning algorithms (e.g., PPO, SAC) and ensures a balanced exploration of critical regions in the real domain. Furthermore, our approach is implemented on the versatile Mujoco MJX platform, and our framework is compatible with a wide range of robotic systems. Experimental results on several robotic manipulation tasks demonstrate that our method significantly reduces the sim-to-real gap, achieving high task performance and generalizability across diverse scenarios of both explicit and implicit environmental uncertainties.

How do humans move? The quest to understand human motion has broad applications in numerous fields, ranging from computer animation and motion synthesis to neuroscience, human prosthetics and rehabilitation. Although advances in reinforcement learning (RL) have produced impressive results in capturing human motion using simplified humanoids, controlling physiologically accurate models of the body remains an open challenge. In this work, we present a model-free motion imitation framework (KINESIS) to advance the understanding of muscle-based motor control. Using a musculoskeletal model of the lower body with 80 muscle actuators and 20 DoF, we demonstrate that KINESIS achieves strong imitation performance on 1.9 hours of motion capture data, is controllable by natural language through pre-trained text-to-motion generative models, and can be fine-tuned to carry out high-level tasks such as target goal reaching. Importantly, KINESIS generates muscle activity patterns that correlate well with human EMG activity. The physiological plausibility makes KINESIS a promising model for tackling challenging problems in human motor control theory, which we highlight by investigating Bernstein's redundancy problem in the context of locomotion. Code, videos and benchmarks will be available at https://github.com/amathislab/Kinesis.