Learning-based manipulation policies typically predict robot actions from sensory observations and hand execution to a separate low-level controller. A new paper from Sungkyunkwan University argues this separation is precisely why rigid-contact tasks go wrong — and proposes a fix that lets the policy decide not just what motion to command, but how the controller should behave while executing it.
The Problem: Policy-Controller Mismatch
In rigid contact, the same motion command can lead to unstable contact, tracking error, excessive loading, or tool damage depending on the low-level controller's behavior. Admittance control tracks motion accurately but can build up dangerous forces; impedance control is safer in contact but sacrifices tracking precision. Most learned policies have no say in which regime executes their commands.
URF: One Framework for Action and Control
The Unified Robot Control-Policy Framework (URF), submitted to IEEE Robotics and Automation Letters (RA-L), connects compliant action prediction with unified impedance-admittance control. Given multimodal observations, URF predicts three things simultaneously:
- A virtual target — where the robot should move
- A stiffness matrix — how rigid each direction should be
- An impedance-admittance switch ratio — when the controller should behave like admittance control for accurate motion tracking versus impedance control for safer rigid contact
Because demonstration data does not provide ground-truth environment stiffness, the team constructed switch-ratio labels from measured contact forces and used them to supervise controller-mode prediction — a practical solution to a key training data gap.
Results: Fewer Broken Tools, Fewer Safety Stops
Across box-flipping and line-pressing tasks, URF achieved higher success rates than an admittance-only baseline (ACP) while eliminating its characteristic failure modes. In box-flipping, ACP broke its end-tool 0.86 seconds after contact, while URF completed the task by lowering stiffness and the switch ratio before contact, then shifting regimes mid-motion as pushing transitioned into lifting. In line-pressing, URF maintained a stable contact force above 5N, while ACP failed to establish stable contact and exhibited large force oscillations before triggering a robot safety stop.
Why It Matters
As VLA models take on contact-rich industrial tasks — from insertion and assembly to surface finishing — the interface between high-level policies and low-level control is becoming a reliability bottleneck. URF's results suggest contact-aware policies benefit from predicting not only compliant actions but also the controller behavior used to execute them, pointing toward tighter policy-control co-design for factory-grade manipulation.
The project page is available at jiyou384.github.io/urf_project_page.
