SAFI teaches VLA robots to stop safely mid-task, cutting takeover success to 1.3%
DM2-ND · hf · 2026-11-21
- Problem: VLA models that refuse a task at the start can still complete it if taken over mid-execution; even an unfinished task may end with the robot dropping a held object — non-completion isn't a complete safety measure.
- Method: State-Aware Fallback Imitation (SAFI) builds fallback targets (stationary arm, consistent gripper hold) for prohibited-task demos, then jointly trains them with retained-task actions so the robot learns an explicit alternative response at intermediate states.
- Results: With OpenVLA-OFT on LIBERO/BadLIBERO, SAFI cuts aggregate target success after takeover to 1.3% with no physical incidents or action saturation, while preserving high retained-task success. Versus RMU and refusal fine-tuning baselines, only SAFI holds the object instead of dropping it.
More from Embodied
- Open-source humanoid Asimov starts wheeled variant on day 389 of development — freelerobot · 2026-11-21
- USC's Heave-Ho trains Unitree G1 humanoids to collaborate, scales 2 to 10 robots — CyberRobooo · 2026-11-21
- Wanting a full self-driving Tesla RV: futuristic Cybertruck-style camper concept — PaulYacoubian · 2026-11-21
- PhysLDM reconstructs deformable trajectories at 2.48mm precision with 78x token compression — jiqizhixin · 2026-11-21
- Open-source autonomous driving book's Chapter 4 compares Pure Pursuit, Stanley, rear-wheel feedback — 4310sy · 2026-11-21
- Berkeley open-sources Specialist: 3D-printed gripper for robot data, runs on real arms — SongShuran · 2026-11-21