RLBench

RLBench is a challenging benchmark and learning environment for robot manipulation introduced by researchers at Imperial College London and Dyson Robot Learning Lab. It features 100 completely unique, hand-designed tasks that span a wide range of difficulty, from simple reaching and door opening to complex multi-stage tasks like opening an oven and placing a tray inside. Built on the CoppeliaSim physics engine with MuJoCo integration, RLBench provides both proprioceptive and visual observations including RGB, depth, and segmentation masks from an over-the-shoulder stereo camera and an eye-in-hand monocular camera. This rich observation space makes it suitable for vision-based manipulation research and sim-to-real transfer. A key feature of RLBench is its procedurally generated task variations. Each task can produce infinite unique episodes through randomization of object positions, orientations, colors, and sizes. This enables robust evaluation of generalization capability. The benchmark also provides motion-planned demonstration trajectories for all tasks, supporting imitation learning and behavior cloning approaches. RLBench has become one of the most widely used manipulation benchmarks, with tasks organized by difficulty tiers. It supports multiple research areas: reinforcement learning (RL, offline RL), imitation learning (behavior cloning, inverse RL), and multi-task learning. The standardized evaluation protocol has been adopted by numerous papers for benchmarking visuomotor policies. With 1,786 GitHub stars, RLBench has an active community and extensive documentation. It integrates with popular RL frameworks like PyTorch and supports customizable task creation. Its successor RLBench 2.0 extends the task suite and adds new capabilities for evaluating generalization and sim-to-real transfer.