RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation

3 min read

RoboTwin 2.0 generates bimanual manipulation data in simulation and benchmarks policies on it. A code agent writes a task program against a skill API from a language description. The program runs 10 times in simulation, a VLM observer watches the runs frame by frame and says which step failed and why, and the code agent repairs the program until it passes a success threshold or five refinements have gone by. Data is then collected under domain randomisation on five axes: distractor clutter, background texture, lighting, table height and instruction wording. The object library (RoboTwin-OD) has 731 objects in 147 categories, each annotated with grasp, placement and functional points, and the benchmark has 50 dual-arm tasks on five embodiments.

Findings

  • Code generation on 10 tasks: average success 62.1% one-shot, 66.7% with execution-log feedback and 71.3% with VLM feedback. RoboTwin 1.0 went from 47.4% to 63.9% under the same three settings.
  • Embodiment-aware grasp candidates raise automated collection success by 8.3% on average over five arms. The gain sits on the 6-DoF arms (Piper +22.7%, Aloha-AgileX +13.7%); the 7-DoF Franka and UR5 barely change.
  • Pretraining on 9,600 randomised trajectories from 32 tasks lifts average success on five unseen tasks, tested under randomisation, from 18.8% to 24.8% for RDT and from 22.5% to 29.1% for Pi0. Fine-tuning on clean data gives almost nothing (RDT drops to 14.6%).
  • Real robot (COBOT-Magic, four tasks, RDT): adding 1,000 randomised synthetic trajectories to 10 clean real demonstrations raises success by 13.5 to 33.0 points across the four test settings. Synthetic data alone gains 21.0 and 20.5 points on the two unseen-background settings.
  • 50-task benchmark, 50 clean demonstrations per task, clean versus randomised test: Pi0 46.4% to 16.3%, RDT 34.5% to 13.7%, ACT 29.7% to 1.7%, DP 28.0% to 0.6%, DP3 55.2% to 5.0%.

Limits

  • The five axes are visual and layout changes, plus up to 1 cm of camera pose jitter in the real-world runs. Mass, friction and control latency, which the related work lists as standard randomisation targets, are not among them.
  • Distractors that look like task objects are excluded on purpose to keep scenes “unambiguous”.
  • One tabletop, two arms, no mobile base, no other robots, no sensor faults.
  • The authors admit DP3’s strong clean score partly comes from perfect point clouds and clean segmentation in simulation.
  • The VLM observer’s diagnoses are used but not themselves scored.
  • Future work named by the authors: real-world deployment and multi-object task complexity.

Relevance

The clean-to-randomised drop (Pi0 from 46.4% to 16.3%) shows how far visual shift alone moves a policy, and nothing here models a sensor that fails or a teammate that lies. The VLM observer is a second model checking the first, the same pattern as the segmenter and reasoner cross-check in GoGet. See embodied-ai-benchmarks, foundation-models-for-robotics and robust robotics.