PhaForce

IROS 2026 / Contact-Rich Manipulation

PhaForce: Phase-Scheduled Visual-Force Policy Learning with Slow Planning and Fast Correction for Contact-Rich Manipulation

Mingxin Wang1 Zhirun Yue1,3 Renhao Lu1,2 Yizhe Li2 Zihan Wang1 Guoping Pan2 Kangkang Dong1 Jun Cheng3 Yi Cheng2 Houde Liu1,*

1Tsinghua University

2Zerith Robotics

3Shenzhen Institutes of Advanced Technology

*Corresponding author: Houde Liu

arXiv PDF Dataset Videos BibTeX
86% Average success rate
+40 pp Improvement over baselines
24 Hz Fast contact correction
5 Tasks Real robot evaluation

Overview

Force feedback is useful only when the policy knows how to schedule it.

Contact-rich manipulation requires vision-dominant task semantics and rapid reactions to force/torque transients. Generative visuomotor policies often update at low frequency because of inference latency and action chunking, so short-horizon contact events can be underused as closed-loop feedback.

PhaForce introduces an explicit contact/phase schedule that coordinates low-rate chunk-level planning with high-rate residual correction. The schedule decides when force should be trusted, how much it should influence planning, and where corrections should be applied during execution.

PhaForce overview for contact-rich manipulation

Method

Slow planning and fast correction share one contact-aware phase schedule.

01

Contact-Aware Phase Predictor

CAP estimates contact probability and task phase belief, turning force history into an interpretable schedule signal for downstream planning and correction.

02

Slow Diffusion Planner

The planner performs dual-gated visual-force fusion with orthogonal residual injection, preserving vision semantics while conditioning action chunks on useful force cues.

03

Fast Residual Corrector

The control-rate corrector routes residual actions through phase-dependent corrective subspaces for within-chunk micro-adjustments.

PhaForce architecture with CAP, Slow planner, and Fast corrector

Experiments

Consistent gains on real robot contact-rich tasks.

Success rate (SR, %) across real-robot tasks.
Method Charger USB Drawer Wipe-ID Wipe-OOD Avg
DP 20 15 60 95 0 38
DP (force-concat) 20 20 50 85 0 35
RDP 50 55 65 85 75 66
PhaForce (ours) 80 85 85 95 85 86
Real robot task success rates
Success rate comparisons across insertion, wiping, and pulling tasks.
Force regulation analysis
Force regulation improves contact quality during physical interaction.
Failure analysis for contact-rich manipulation
Phase-routed corrections reduce typical jamming and misalignment failures.

Videos

Qualitative demonstrations across contact modes.

Wiping / ID Stable contact tracking on the training-style surface.
Wiping / OOD Robust adaptation under shifted surface geometry.
Plug Insertion Corrective response during constrained insertion.
USB Insertion Search and alignment from contact feedback.
Drawer Pulling Force-aware transition through grasp, contact, and pull.

Citation

BibTeX

@misc{wang2026phaforce,
  title={PhaForce: Phase-Scheduled Visual-Force Policy Learning with Slow Planning and Fast Correction for Contact-Rich Manipulation},
  author={Wang, Mingxin and Yue, Zhirun and Lu, Renhao and Li, Yizhe and Wang, Zihan and Pan, Guoping and Dong, Kangkang and Cheng, Jun and Cheng, Yi and Liu, Houde},
  year={2026},
  eprint={2603.08342},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  doi={10.48550/arXiv.2603.08342}
}