OpenAI Fetch-style robot tasks, rebuilt for Gymnasium and a Panda arm

What it does

panda-gym is a collection of six manipulation environments—reach, push, slide, pick-and-place, stack, and flip—built around a simulated Franka Emika Panda robot. It wraps PyBullet physics in the Gymnasium API, letting standard RL libraries drive the arm without ever endangering physical hardware. The tasks are explicitly modeled after OpenAI’s older Fetch environments, ported to the current gymnasium standard.

The interesting bit

Rather than inventing new challenges, the project carefully translates a well-known robotics benchmark suite to a widely used open-source arm and a modern API. That makes it a convenient drop-in replacement when you are reproducing older papers or benchmarking new algorithms against a familiar baseline.

Key highlights

• Six standard manipulation tasks: PandaReach-v3, PandaPush-v3, PandaSlide-v3, PandaPickAndPlace-v3, PandaStack-v3, and PandaFlip-v3.
• Built on PyBullet physics and the gymnasium interface, compatible with standard RL training loops.
• Pre-trained baseline agents and training results available through rl-baselines3-zoo and the Hugging Face Hub.
• Published academic paper (arXiv:2106.13687) and actively maintained with CI, code coverage, and documentation.

Caveats

• The README notes the environments are “widely inspired” by OpenAI Fetch rather than direct re-implementations, so exact behavioral parity with the originals is unclear.

Verdict

Worth a look if you need a lightweight, off-the-shelf manipulation benchmark for a standard RL pipeline. Skip it if you are looking for diverse robot morphologies or high-fidelity physics beyond PyBullet’s scope.