Assistive Gym

Physics simulation framework for training assistive robots in caregiving tasks

License

MIT

Overview

Assistive Gym is a physics-based simulation framework, built on PyBullet and OpenAI Gym, for training robots to help people with daily tasks such as feeding, drinking, bathing and dressing. It pairs six assistive tasks with six commercial robots and builds human preferences into its reward functions. It includes pretrained baseline policies for reinforcement learning.

Assistive Gym is a physics-based simulation framework built on PyBullet and integrated with OpenAI Gym, designed specifically for research in physical human-robot interaction and robotic assistance. It provides 72 ready-to-use simulation environments combining six assistive tasks (feeding, drinking, itch scratching, bed bathing, dressing, and arm manipulation) with six commercial robots (PR2, Jaco, Baxter, Sawyer, Stretch, and Panda) in both static and active human configurations.

The framework features anatomically realistic human models with 40 actuated joints and pose-dependent joint limits derived from biomechanical research, supporting both male and female body configurations. It incorporates human preference modeling directly into reward functions, enabling robots to learn assistance strategies that respect a person's comfort and autonomy. Version 1.0 added mobile manipulation, multi-robot control, iGibson integration for photorealistic home environments, and SMPL-X human mesh models for realistic body representation.

Assistive Gym is fully compatible with standard reinforcement learning algorithms through the OpenAI Gym interface, includes pretrained baseline policies for all robot-task combinations, and supports cloud-based development through Google Colab notebooks. Researchers can create custom assistive environments, extend existing tasks, and benchmark new approaches against provided baselines.

Why it matters

As populations age globally, the demand for caregiving far outpaces the available workforce. Assistive robotics holds the potential to bridge this gap, but developing robots that physically interact with people is inherently risky. Assistive Gym provides a safe, repeatable simulation environment where researchers can train and evaluate assistive robot policies without putting real people at risk during the development process.

The framework has established itself as a foundational tool in the assistive robotics research community, published at IEEE ICRA 2020 and cited extensively. By standardizing the simulation of common caregiving tasks and providing benchmark results across multiple commercial robots, it accelerates progress in a field where real-world experimentation is expensive, slow, and ethically constrained. Its open-source nature and Google Colab support lower the barrier to entry for new researchers entering this critical domain.

Problem it solves

Developing robots that physically assist people with daily activities presents a fundamental chicken-and-egg problem: robots need extensive training to interact safely with humans, but training on real people poses unacceptable safety risks. Physical human-robot interaction experiments are also expensive, time-consuming, and difficult to reproduce across research groups.

Assistive Gym solves this by providing a realistic physics simulation where robots can learn through millions of trial-and-error interactions without any risk to human subjects. The framework models not just the physics of contact, but also human preferences and physical capabilities, ensuring that trained policies account for the comfort and autonomy of the person being assisted. This enables researchers worldwide to develop, test, and compare assistive strategies on standardized benchmarks before any real-world deployment.

Key takeaways

• Simulates six daily living tasks (feeding, drinking, bathing, dressing, scratching, arm manipulation) across six commercial robots with 72 total environment combinations
• Integrates human preference modeling into reward functions so robots learn to assist in ways that respect comfort and autonomy
• Features anatomically realistic human models with 40 actuated joints and biomechanically accurate joint limits
• Compatible with standard reinforcement learning through OpenAI Gym interface, with pretrained baseline policies included
• Supports photorealistic home environments via iGibson integration and realistic body representation through SMPL-X human meshes

Who it is for

• Robotics researchers working on physical human-robot interaction and assistive technology
• Reinforcement learning practitioners applying RL to real-world caregiving and manipulation tasks
• Healthcare technology developers prototyping robotic assistance for elderly and disabled care
• Graduate students studying assistive robotics, rehabilitation engineering, or embodied AI
• Robot manufacturers evaluating commercial platforms for assistive applications

Limitations

• Simulation-to-reality gap means policies trained in Assistive Gym may not transfer directly to physical robots without additional fine-tuning
• Human models use simplified joint limits and body dynamics that cannot capture the full variability of real patients with different disabilities
• Requires Python 3.6 or 3.7 for best compatibility, with limited support for newer Python versions
• Does not model soft-body deformations for tasks like dressing and bathing where cloth and skin physics are important
• Last significant update was v1.0 and the project has limited recent development activity compared to newer simulation frameworks