Real-World RL Suite
Google Research's benchmarks for reinforcement learning that works on physical robots. Safety constraints, noise, delays -- the messy realities of deployment.
Overview
The Real-World RL Suite addresses a critical gap in RL research: most benchmarks (Atari, MuJoCo, DMControl) assume perfect sensing, instant actuation, and unlimited exploration. Real robots have none of these luxuries. This suite provides environments and evaluation protocols that incorporate safety constraints (don't break the robot), partial observability (sensors fail), observation noise (real cameras are noisy), action delays (communication latency), and non-stationarity (environments change).
It is built on top of the DeepMind Control Suite but adds real-world challenges as configurable wrappers, making it straightforward to evaluate how robust any RL algorithm is to deployment conditions.
Access
Related datasets
- ManiSkill2 -- manipulation benchmark
- Habitat -- embodied AI environments
- RoboCasa -- kitchen simulation