RLBench Robot Learning Benchmark
Provider：Imperial College London (Dyson Robotics Lab)
RLBench offers 100 hand-designed manipulation tasks with unlimited motion-planned demonstrations, standard protocol 18 tasks/249 variations on a Franka Panda in CoppeliaSim.
At a glance
- Readiness
- L2 · Semantic & material ready
- Rationale
- Manipulation L2: 100 hand-designed manipulation tasks, each with procedural variations and an unlimited supply of motion-planned demonstrations in the CoppeliaSim physics scene; grasping uses parent-child attachment rather than frictional contact, so it is not L3.
- Scale
- 100 hand-designed tasks with procedural variations; the standard evaluation protocol uses 18 tasks and 249 variations with 100 demonstrations per task (The RLBench paper (arXiv 1909.12271, RA-L 2020) lists 100 tasks; the 18-task/249-variation protocol used since PerAct comes from follow-up work. Checked 2026-10-25.)
- License
- Research use only — A custom Imperial College academic license (BSD-style with mandatory citation of the RLBench paper) permits research use and incorporation into products, but check the LICENSE before redistribution or commercial use.
- Access
- Open
- Formats
- other
- Runtimes
- pybullet · other
- Physics properties
- Collision geometryJoint structureSensor attributes
- Domain
- Other
- Published
- 2026-10-25
- Updated
- 2026-10-25
Why it is listed
RLBench solved a reproducibility problem in manipulation research: instead of each paper inventing its own tasks, it supplies a shared scene, a shared task library and a shared observation interface, so perception-to-action pipelines can be compared under identical conditions.
Readiness assessment
One hundred hand-designed tasks span simple target reaching to multi-stage activities such as opening an oven and placing a tray inside. Each task ships waypoints that a motion planner turns into an effectively unlimited supply of expert demonstrations, which is what made demonstration-based learning practical at this scale.
Practical notes
The environment runs on CoppeliaSim v4.1.0 via PyRep with a Franka Panda and parallel gripper; arms can be swapped, UR5 support arrived in v1.1.0. Grasping is deliberately simplified by attaching the object to the gripper tip, which speeds up simulation but means contact physics is not realistic.
Disclaimer: license and usage terms follow the official source — verify them yourself before use. This site indexes and links only; it does not host or mirror any asset file.
License and usage terms are those of the official source — verify before use. No asset file is hosted or mirrored on this site.