Force-informed robot learning

FoLD: Force-Informed Learning for Dexterous Articulated Object Manipulation

Haowei Shen¹Tingai Li¹Yumeng Liu^1,*Wenyuan Guang²Xuanze Yang¹Qing Fang¹Kai Xu^2,3Ligang Liu¹Ruizhen Hu^4,*

¹University of Science and Technology of China ²Institute of AI for Industries, Chinese Academy of Sciences ³Jiangsu Key Laboratory of AI for Industries ⁴Shenzhen University

arXiv Watch overview

Force-Informed Learning for Dexterous Articulated Object Manipulation turns object-motion deviations into compensatory force guidance, enabling diverse robotic hands to refine retargeted motions into physically effective interactions.

Overview

FoLD in three minutes

From human demonstration and kinematic retargeting to compensatory force guidance, simulation learning, and real-robot deployment.

Abstract

Force guidance bridges the embodiment gap.

Transferring human demonstrations to dexterous robots remains challenging because differences in hand morphology and contact dynamics often cause retargeted motions to fail at producing the intended object behavior. We present FoLD, a framework for learning dexterous manipulation of articulated objects through explicit force guidance. FoLD computes compensatory force fields from human demonstrations together with the robot’s current interaction state, yielding a force prior that promotes the demonstrated object motion. This force prior informs a residual policy that adapts retargeted hand motions to the contact requirements of the task. We evaluate FoLD on a public benchmark for articulated object manipulation, where it consistently outperforms state-of-the-art baselines across tasks and embodiments. We further validate FoLD on real dexterous robot platforms, demonstrating successful transfer of human manipulation skills to robot execution.

Method

Correct the interaction, not just the pose.

FoLD measures object-motion deviation, estimates a compensatory force field, and uses it as a prior for learning corrective residual actions.

Initial imitation

Retarget the human motion to a robot hand while the object is driven along the ground-truth physical trajectory (GT object trajectory). Even with this GT motion, the resulting hand–object contact can remain imprecise.

Compensatory force field

Convert the discrepancy between desired and realized object motion into a smooth spatial force prior that highlights where corrective interaction is needed.

Force-informed refinement

Learn residual hand actions under force guidance so that robot contacts produce the intended articulated motion instead of merely matching hand pose.

Qualitative results

Cross-embodiment comparison with different baselines.

We compare human demonstrations, FoLD, and the corresponding imitation baseline on articulated-object tasks.

FoLD vs. DexMachinaSix examples across Inspire, Allegro, XHand, and Schunk hands

Notebook

Box

Ketchup

Waffle Iron

Ketchup · Allegro

Mixer · XHand

FoLD as a CHORD enhancementHuman · CHORD + FoLD · CHORD

Box

Notebook

Quantitative results

Better object motion across embodiments.

Across four robotic hands, FoLD raises overall success rate from 0.606 to 0.772 while reducing position, rotation, and joint errors.

Comparison with DexMachina

Best result for each hand is bold.

Hand	Method	SR ↑	Contact mean ↑	Pos. (cm) ↓	Rot. (°) ↓	Joint (°) ↓
Inspire	DexMachina	0.511	0.154	52.27	42.65	34.26
Inspire	FoLD	0.700	0.248	18.19	23.28	20.20
Allegro	DexMachina	0.596	0.043	20.07	49.38	22.51
Allegro	FoLD	0.750	0.061	23.62	40.51	12.18
XHand	DexMachina	0.724	0.113	39.55	48.34	24.88
XHand	FoLD	0.697	0.116	31.86	43.76	20.24
Schunk	DexMachina	0.588	0.034	34.47	45.66	44.49
Schunk	FoLD	0.968	0.101	0.91	6.52	5.37
Overall	DexMachina	0.606	0.088	36.67	46.54	31.06
Overall	FoLD	0.772	0.133	19.30	29.33	14.83

FoLD added to CHORD

Sharpa-Wave hand; mean ± sample standard deviation.

Method	SR ↑	CWS mean ↑	Pos. (cm) ↓	Rot. (°) ↓	Joint (°) ↓
CHORD	0.392 ± 0.315	0.351 ± 0.242	32.90 ± 29.47	43.58 ± 38.03	21.92 ± 23.45
CHORD + FoLD	0.397 ± 0.314	0.344 ± 0.225	20.19 ± 17.92	33.60 ± 30.40	22.17 ± 24.00

Adding FoLD substantially reduces object position and rotation errors while maintaining a similar success rate and contact consistency.

Real-robot deployment

From simulation to physical interaction.

FoLD transfers a range of articulated manipulation skills to a bimanual dexterous robot platform.

Laptop

Headphones

Scissors

Failure casesView two representative physical failures

Citation

The paper is now available on arXiv.

Read the manuscript on arXiv:2609.33551. The BibTeX entry and source code will be added when the public release is ready.