Title: Assistax: A Multi-Agent Hardware-Accelerated Reinforcement Learning Benchmark for Assistive Robotics

Abstract:

The advancement of reinforcement learning (RL) algorithms has historically been propelled by ambitious challenge tasks and standardized benchmarks. Video games have long dominated this landscape, favored for their ability to present relevant difficulties, low computational costs, and intuitive understandability. While titles like Go and Atari have catalyzed numerous breakthroughs, their applicability often fails to translate directly to real-world embodied scenarios. To address the necessity of diversifying RL benchmarks and to tackle the complexities inherent in embodied interaction, we present Assistax: an open-source benchmark specifically designed for challenges found in assistive robotics.

Assistax leverages JAX’s hardware acceleration to significantly enhance training speeds within physics-based simulations. Regarding open-loop wall-clock time, the benchmark achieves performance up to $370\times$ faster than CPU-based alternatives when vectorizing training runs. Assistax models the dynamic between an assistive robot and an active human patient through multi-agent RL, training a diverse population of partner agents. This setup allows for the testing of an embodied robotic agent’s zero-shot coordination capabilities. Through extensive evaluations and hyperparameter tuning across popular continuous control RL and MARL algorithms, we provide reliable baselines, establishing Assistax as a practical tool for advancing RL research in the field of assistive robotics. The code is available at: https://github.com/assistive-autonomy/assistax.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC