A gym for robots that need both hands

What it does RoboTwin is a simulation platform built on IsaacLab for dual-arm robotic manipulation. It generates synthetic demonstration trajectories and provides standardized tasks to train and evaluate policies like ACT, DP3, RDT, and various VLA models. The project also maintains a public leaderboard and has released over 100,000 pre-collected trajectories, though it nudges users toward generating their own to match specific task and robot configurations.

The interesting bit The “strong domain randomization” pitch is the core bet: instead of chasing photorealism, it randomizes textures, lighting, and object placements aggressively so policies trained in simulation transfer better to messy reality. The platform also tracks cross-embodiment performance — whether a policy trained on one robot arm geometry survives on another — which is where a lot of sim-to-real research quietly falls apart.

Key highlights

• Supports 50+ bimanual tasks with configurable difficulty and language conditioning
• Ships integration branches for IsaacLab-Arena, RLinf, and StarVLA
• Public leaderboard tracks visual robustness, language diversity, and multi-task generalization
• CVPR 2025 Highlight paper; active challenge track at CVPR 2025 MEIS Workshop
• Active maintenance: bug fixes for Piper wrist, DP3 eval, ACT deployment posted in recent months

Caveats

• The 2.0 paper is still under review as of the README date; some claims remain unvetted
• Several listed policy integrations (G3Flow, HybridVLA, SmolVLA, AVR, UniVLA) are marked TODO
• Pre-collected dataset is large but explicitly described as less configurable than user-generated data

Verdict Worth a look for robotics researchers working on imitation learning or VLA models who need a bimanual benchmark with less hardware overhead. Less useful if your problem is single-arm or you already have a reliable real-world data pipeline you prefer.