FORCA: A Force-Aware Coding Agent for Contact-Rich Robot Manipulation

At a glance

Problem: coding agents mostly rely on vision and proprioception, which leaves geometry, mass, friction and contact uncertain in contact-rich manipulation. Source
Idea: the agent writes a complete Python task program that probes contact, runs force feedback loops, and is revised from recorded force traces after each rollout. Source
Evidence: on hand-held vase wiping, FORCA removed 93% of the marker versus 21% for the vision-only agent, averaged over 10 trials. Source

Evaluation scope Every comparison is against a vision-only agent that uses the same foundation model, instructions and basic skills but receives no force data in program generation, revision or execution. Source

How it works

The agent writes the force-dependent logic for each task from the wrist wrench exposed by the robot interface, for example when an approach should stop or how motion responds to a force error. Source
During a rollout the program can probe a surface and use the robot pose at contact to update its estimate of surface height or grasp position before acting. Source
For sustained contact, a host-side 100 Hz loop reads the wrench and adjusts Cartesian velocity, for example pausing forward motion until contact is restored. Source
After each rollout the agent inspects visual observations with force and robot-state traces, can write analysis code, and revises parameters, contact conditions or control logic. Source

Illustrative revision for cucumber peeling: the agent links peeling gaps to lost contact in the force record and adds a wait-for-contact condition. The paper notes that force traces and skin patches are idealized and the code simplified. Original figure

Evidence & results

Over 10 trials, drawer and door-handle success rose from 40% and 10% for the vision-only agent to 100% for FORCA, with lower grasp translation and rotation errors. Evidence
Peeling a hand-held cucumber over 10 trials, contact loss fell from 40.6% to 2.2% and the peeling ratio rose from 37.4% to 81.1% compared with the vision-only agent. Evidence
After developing a knocking program for a cube on a whiteboard, each of three new object or surface configurations reached the target band within a few extra iterations. Evidence

Cucumber peeling rollouts with the hand-held setup, peeled strips from one representative pass, execution frames and normal contact force with contact-loss intervals. Original figure

Limitations & boundaries

Each iteration takes several minutes, mostly agent reasoning, and refinement still depends on real-robot rollouts. Source
Wrist wrench sensing does not reveal where contact occurs or how forces are spread across the fingers. Source
For longer action sequences or tighter force constraints, the agent may fail to produce a working program even after repeated attempts. Source
Wiping and peeling used the full budget of 20 development iterations, with agent reasoning taking 4–10 minutes of each iteration. Source

Takeaway

Our reading: the paper's main point is that contact experience can live in inspectable code, since a program refined with force feedback could be reused and revised when the object or surface changed. Source

Resources

Full text ↗PDF ↗

Figures: the paper's authors. Editorial source and image review: 2026-10-26. Findings are author-reported, not independently reproduced by Robotics Research Hub.

Original abstract

Coding agents can translate language instructions into robot programs. For contact-rich manipulation, a key challenge is to generate and refine control logic that responds to physical interaction. We present FORCA, a coding agent that takes visual observations and wrist force/torque (F/T) measurements as input to write Python programs for contact probing and force-responsive motion. Between executions, the agent inspects force records alongside visual outcomes to diagnose failures and revise both motion and force-feedback control logic. FORCA requires no demonstrations, task-specific training, or simulation. Across five real-robot tasks, we demonstrate its effectiveness in three forms of contact-rich manipulation: improving manipulation accuracy, adapting to changes in object or surface properties, and regulating contact under surface variation and object motion. Compared with a vision-only variant of the same agent, FORCA improves grasp accuracy and task success and reduces contact loss in wiping and peeling.

Citation

@article{Tan2026_261031841,
  title     = {FORCA: A Force-Aware Coding Agent for Contact-Rich Robot Manipulation},
  author    = {Xiao Tan and Yixuan Wang and Zhexi Wang and Wenzhen Yuan},
  year      = {2026},
  eprint    = {2610.31841},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url       = {https://arxiv.org/abs/2610.31841}
}