Researchers at Shanghai Jiao Tong University and the Shanghai Institute of Innovation report that ActSafeGuard kept robot commands within specified movement limits across simulated tasks and small physical trials. They published the method in a September 10 preprint.
The system works with models such as pi0.5, which turn camera images and language instructions into sequences of motor commands. Even after learning from valid demonstrations, these models can generate commands that exceed allowed joint positions or speeds.
ActSafeGuard starts from a command sequence inside those limits. As the model revises the sequence, it scales back any update that would cross a boundary. The same check is included during training, allowing the model to learn how its proposed movements are being constrained.
In four simulated tasks, pi0.5's average completion rate rose from 75.25% to 81.5% when both position and speed limits were enforced. A physical robot also guided a ball along an L-shaped track successfully in five out of five trials, compared with four for the unguarded model.
The limits are still specified by people or derived from demonstrations. The authors identify learning and updating them automatically as future work.
Sources
- Jianming Ma and colleagues: ActSafeGuard, arXiv:2609.11697v1 (September 10, 2026). Methods, Table 1 and real-robot evaluation.
- Physical Intelligence: pi0.5 model paper, arXiv:2504.16054v1 (April 22, 2025). Background on the underlying robot model.
- Illustrative file photograph: Lexington Medical, Inc., manufacturing robot arm (2025), CC0 1.0. An unrelated manufacturing installation, shown at its original 16:9 proportions.
