[CVPR 2024 / CoRL Potential] Force Policy: Mastering the High-Frequency Logic of Contact-Rich Manipulation

TL;DR

Force Policy bridges the gap between the coarse guidance of vision and the high-frequency precision of force feedback. By formalizing a physically-grounded Interaction Frame (IF) and employing a Global-Local architecture, this method allows robots to perform "hard" tasks—like plugging in an EV charger requiring 160N or scraping off stickers—with human-like dexterity and stability.

Problem & Motivation: The "Vision-Force Entanglement" Trap

In human motor control, vision tells us where to Reach, while haptics tell us how to Interact. Most modern AI policies (like VLA models or standard Diffusion Policies) treat force as just another sensor reading—appending it to a visual observation vector.

This leads to two critical failures:

1. Pollution: Noisy force signals during free-space motion confuse the global planner.
2. Latency: Global vision models are too slow (low frequency) to react to the millisecond-level transients when a robot hits a hard surface, leading to "contact instability" or jamming.

The authors argue that we need an explicit Interaction Structure—a way to tell the robot which directions to move in (Position Control) and which directions to exert pressure in (Force Control).

Methodology: The Interaction Frame (IF)

The core innovation is the Interaction Frame (IF). Unlike the classical "Task Frame Formalism" which requires a CAD model of the object, Force Policy recovers this frame directly from interaction data using an energy dissipation perspective.

1. Physical Decoupling

The IF is defined as a local basis where:

• The Z-axis is the dominant wrench (force) direction.
• The X-axis is the intended motion direction.

This allows the robot to execute Hybrid Force-Position Control. In a peg-in-hole task, it provides position control to align the peg and force control to push it in without snapping the tool.

Fig 1: The Global-Local architecture. The global policy (RISE-2) provides context, while the local policy (ResNet+GRU) manages high-speed reactive adjustments.

2. Dual-Policy Asynchronous Scheduler

To solve the latency problem, the system runs two threads. The local force policy runs at 50Hz, while the global vision policy can take its time. They use Dynamic Time Warping (DTW) to align the "chunks" of predicted actions, preventing the jerky "retreating" motions common in high-latency robotic systems.

Experimental Evidence: Smashed Baselines

The researchers tested Force Policy across three aggressive scenarios: Push and Flip, Plug in EV Charger, and Scrape off Sticker.

Key Result: Precise Force Regulation

Standard vision-only models (like RISE-2 or π0.5) failed the insertion and scraping tasks because they simply cannot "feel" the resistance. Meanwhile, monolithic vision-force models (like ForceVLA) exhibited severe oscillations.

Fig 2: Comparison of force profiles. Force Policy (red) closely tracks the human expert's force signature, whereas baselines either oscillate or apply insufficient pressure.

Generalization to the Unseen

Because the IF is based on local physics rather than global visual appearance, the policy generalized effortlessly to:

• Geometric shifts: Flipping cylindrical cups instead of boxes.
• Material shifts: Handling soft sponges vs. hard wood blocks.
• Visual shifts: Novel colors and textures.

Critical Insight & Conclusion

Force Policy demonstrates that we shouldn't just "throw more data" at a monolithic transformer and hope it learns physics. By re-introducing Hybrid Force-Position Control as a structured layer within a modern neural policy, we gain both the generalization of deep learning and the stability of classical control theory.

Limitations: The current model assumes the Interaction Frame origin is at the end-effector. For complex tools or destructive tasks (like cutting), a more sophisticated contact-point estimation would be required.

The Takeaway: The future of dexterous robots lies in "Global Planning + Local Reactivity." Separating these roles is the only way to handle the high-friction, high-uncertainty world of real-world contact.