Online Safety Filter for Deformable Object Manipulation with Horizon Agnostic Neural Operators
Robots can now safely handle fluids and cloth with real-time constraint enforcement.
A new online safety filter for deformable object manipulation enforces explicit task-level safety constraints in real time by minimally modifying any nominal control policy. The method combines a horizon-agnostic neural operator that learns boundary input-output mappings of underlying PDE dynamics with a boundary control barrier function that certifies safety via a lightweight quadratic program. Evaluated on fluid manipulation tasks in FluidLab, the filter improves safe trajectory rates by up to 22% over unfiltered base policies while also reducing steps to reach the safe set.
- Combines a horizon-agnostic neural operator (PDE boundary mapping) with a boundary control barrier function for real-time safety.
- Achieves up to 22% improvement in safe trajectory rates on FluidLab fluid manipulation tasks.
- Reduces number of steps needed to reach safe set compared to reward-shaping baselines.
Why It Matters
Enables safer, constraint-guaranteed robot manipulation of deformable objects, critical for industrial and medical applications.