← All benchmarks

Dual-Arm Sim (task completion)

Agents · 2026-10-30

micro1's Dual-Arm Sim has a general-purpose model drive two simulated Franka Panda arms at an operating-room instrument trolley in NVIDIA Isaac Sim and Isaac Lab. At each of at most 120 decisions it gets an overhead camera image, a wrist camera image per arm, fingertip positions and gripper openings, and returns relative arm moves, gripper commands, target estimates, a status note and a done flag; the arm executes about 40% of each commanded move, which the model is not told. Four tasks (instrument swap, drug-cartridge swap, two-arm sample transfer and surgical-kit packing) run 20 trials each, on 20 seeded layouts shared by every model, except the instrument swap, which uses one fixed layout. The score is the share of a model's 80 trials in which every task condition, including the arm-use rules, held at the end, judged from simulator ground truth.

Top models (higher is better)