RoboChemGym: A Protocol-Driven Generative Simulation Framework for Long-Horizon Chemical Manipulation

Can generative simulation be extended to automatically produce protocol-aligned, long-horizon chemical manipulation demonstrations and support hierarchical evaluation of embodied agents?

Wet-lab experiments are expensive, slow, and safety-sensitive, but training embodied lab robots is blocked by scarce real-world manipulation data. Existing simulation pipelines rely on manual setup or teleoperation; generative simulators tend to target short, loosely structured everyday tasks, and lab-specific simulators often cover isolated operations rather than complete protocols. Chemical work adds strict step ordering, fine-grained actions, multi-object layouts, reachability limits, collisions, and occlusion, which current methods do not jointly handle.

Earlier work includes rigid/fluid-capable simulators such as MuJoCo, PyBullet, SAPIEN, and Isaac-based environments, and generative simulators such as GenSim, RoboGen, and ReGen. Dedicated lab environments still depend on manual configuration and post-hoc success checks. These methods are not designed to jointly refine action parameters and scene layouts against execution failures, so long protocol sequences break down.

RoboChemGym starts from a natural-language protocol and uses Claude 3 Opus to extract scene descriptions and action sequences. A scene-construction engine places assets under workspace, collision, and occlusion constraints; an action planner composes eight parameterized primitives—Pick, Place, Pour, Press, Open, Close, Shake, Stir. A closed-loop self-improving stage monitors execution continuously, then runs two optimization loops: an inner action loop adjusts primitive parameters, and an outer scene loop repositions objects when action fixes fail. Action and scene memory store prior diagnoses and corrections. Lab-centric data scaling randomizes object pose, background clutter, workbench texture, and illumination. The benchmark spans 20 protocol-derived tasks at 3, 7, and 11 atomic actions, with both step-wise and full-sequence success metrics.

Generation: clean success is 89.2% for 3-step, 73.4% for 7-step, and 52.3% for 11-step tasks; randomized settings drop only about 3.8 points on average to 85.6%, 70.3%, and 47.5%. Removing action memory reduces clean success by 13.4–24.9 points, including 52.3% to 27.4% on Level 3; removing scene memory reduces it by 15.6–20.0 points. Real-robot replay of generated trajectories reaches 85.7%, 69.2%, and 54.6% without reset actions, and 87.3%, 79.7%, and 70.5% with pick/place resets. For policies trained on 200 expert trajectories per task, π0+random reaches 68.5% full-sequence on Level 1, 28.4% on Level 2, and 1.7% on Level 3; π0 zero-shot reaches 0.8%, 0%, and 0%. ACT+random reaches 35.2% on Level 1 and 12.4% on Level 2; Diffusion Policy+random reaches 25.6% and 8.5%. Real-robot: combining 1,000 synthetic trajectories with 10 real demonstrations on three Level-1 tasks improved π0 average success from 14.5% to 37.8% across four configurations, a 23.3-point gain; synthetic-only zero-shot improved by 4.5–6.0 points over the real-only baseline.

Authors state that real-world validation is so far trajectory replay, not fully autonomous execution, and the framework supports only the Franka Panda arm. Simulation results also show very low full-sequence policy success on 11-step tasks, with π0+random at 1.7%, so true long-horizon autonomy remains unsolved. Real-robot policy evaluation covers only three Level-1 tasks with a small real-demonstration budget. The paper does not state whether code or assets will be released.