FoLD: Force-Informed Learning for Dexterous Articulated Object Manipulation

Haowei Shen; Tingai Li; Yumeng Liu; Wenyuan Guang; Xuanze Yang; Qing Fang; Kai Xu; Ligang Liu

cs.ROSep 27, 2026

FoLD: Force-Informed Learning for Dexterous Articulated Object Manipulation

Authors: Haowei Shen, Tingai Li, Yumeng Liu, Wenyuan Guang, Xuanze Yang, Qing Fang, Kai Xu, Ligang Liu, +1 more

Organizations: University of Science and Technology of China · Institute of AI for Industries, Chinese Academy of Sciences · Jiangsu Key Laboratory of AI for Industries · Shenzhen University

Abstract

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 \textbf{FoLD}, a framework for learning dexterous manipulation of articulated objects through explicit force guidance. FoLD compute 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. Here is the link of our project page: https://gghgghgghgg.github.io/FoLD-project-page/.

Figures & tables

Fig. 2: Overview of FoLD. The imitator provides reference robot actions from a human demonstration. As the robot executes these actions, FoLD estimates a compensatory force field from the resulting object-motion deviations and the current contact map. The field is encoded into a latent representation that informs a residual policy, which corrects the reference actions to bring the object motion closer to the demonstration.

Fig. 3: Qualitative results on mixer, ketchup-bottle, and notebook manipulation (top to bottom). From left to right: human demonstrations, DexMachina, hand-only imitation, and FoLD, with two snapshots per group. Hand-only imitation uses only our pretrained imitator’s hand actions, with the object’s root pose and articulation fixed to the demonstrated state at each frame.

Hand	Method	SR $\uparrow$	Contact Mean $\uparrow$	Pos. (cm) $\downarrow$	Rot. ( ∘ ) $\downarrow$	Joint ( ∘ ) $\downarrow$
Inspire	DexMachina	0.511	0.154	52.27	42.65	34.26
Inspire	Ours	0.700	0.248	18.19	23.28	20.20
[0.5pt/5pt] Allegro	DexMachina	0.596	0.043	20.07	49.38	22.51
[0.5pt/5pt] Allegro	Ours	0.750	0.061	23.62	40.51	12.18
[0.5pt/5pt] XHand	DexMachina	0.724	0.113	39.55	48.34	24.88
[0.5pt/5pt] XHand	Ours	0.697	0.116	31.86	43.76	20.24

TABLE I: Comparison of DexMachina and FoLD across robotic hands. Both methods use the same 27 paired hand–task evaluations (aggregation details in Appendix D-B ). Best results for each hand are shown in bold.

Fig. 4: Bimanual manipulation across hand embodiments. Human configurations appear on the left, followed by robot configurations learned with FoLD.

Fig. 5: Additional FoLD manipulation examples. Each group presents three snapshots showing coordinated support and manipulation of articulated object parts.

Method	SR $\uparrow$	CWS Mean $\uparrow$	Pos. (cm) $\downarrow$	Rot. ( ∘ ) $\downarrow$	Joint ( ∘ ) $\downarrow$
CHORD	$0.392{\pm}0.315$	$\mathbf{0.351{\pm}0.242}$	$32.90{\pm}29.47$	$43.58{\pm}38.03$	$\mathbf{21.92{\pm}23.45}$
CHORD + FoLD	$\mathbf{0.397{\pm}0.314}$	$0.344{\pm}0.225$	$\mathbf{20.19{\pm}17.92}$	$\mathbf{33.60{\pm}30.40}$	$22.17{\pm}24.00$

TABLE II: Comparison of CHORD and CHORD + FoLD on Sharpa-Wave dexterous hand. Values are mean

\pm

sample standard deviation.

Fig. 6: Human demonstrations, CHORD, and CHORD + FoLD on box, notebook, and laptop manipulation (top to bottom). These snapshots illustrate hand configurations relative to the articulated object parts.

Residual policy	VOC curriculum	Force-field input	SR $\uparrow$	Contact Mean $\uparrow$	Pos. (cm) $\downarrow$	Rot. ( ∘ ) $\downarrow$	Joint ( ∘ ) $\downarrow$
✓	✗	✗	0.848	0.300	1.04	11.33	11.99
✓	✓	✗	1.000	0.275	1.34	5.55	13.30
✓	✓	✓	1.000	0.351	0.84	7.69	13.76

TABLE III: Component ablation on Inspire. Checkmarks and crosses indicate enabled and disabled components, respectively. All rows use the same frozen imitator; the last row is FoLD as evaluated in the main comparison. Bold indicates the best result.

Fig. 7: Real-robot demonstrations of laptop closing (upper two rows) and headphone folding (lower two rows). Each task pairs the human demonstration with open-loop replay of the FoLD-generated trajectory on two FR3 arms with Inspire hands.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
Position proportional gain $k_{p}$	200
Position derivative gain $k_{d}$	20
Rotation proportional gain $k_{R}$	20
Rotation derivative gain $k_{\omega}$	2
Force command scale	$25\,\mathrm{N}$
Torque command scale	$3\,\mathrm{N\,m}$

Appendix

TABLE IV: Online compensatory force-field construction parameters.

Parameter	Value
Parallel environments	4096
Rollout horizon	16
Learning rate	$3\times 10^{-4}$
PPO mini-epochs	5
Minibatch size	32768
Discount factor $\gamma$	0.99

Appendix

TABLE V: PPO and simulation configuration used in the DexMachina and FoLD experiments.

Term	DexMachina	FoLD
Object tracking	1.0	1.0
Hand-pose imitation	0.3	0.3
Frame-to-frame motion consistency	0.0	0.1
Joint-reference tracking	0.3	0.3
Position-based contact tracking	3.0	3.0
High-contact-force penalty	0.1	0.1

Appendix

TABLE VI: Reward coefficients for the DexMachina baseline and FoLD stage-2 residual-policy training.

Task	Training identifier	ARCTIC sequence	Frame range	Length
Ketchup-100	ketchup-30-130-s01-use_01	s01/use_01	$[30,130)$	100
Box-200	box-30-230-s01-use_01	s01/use_01	$[30,230)$	200
Mixer-170	mixer-30-200-s01-use_01	s01/use_01	$[30,200)$	170
Ketchup-300	ketchup-30-330-s01-use_01	s01/use_01	$[30,330)$	300
Mixer-300	mixer-30-330-s01-use_01	s01/use_01	$[30,330)$	300
Notebook-300	notebook-30-330-s02-use_01	s02/use_01	$[30,330)$	300

Appendix

TABLE VII: ARCTIC demonstrations used in the four-hand DexMachina comparison. Frame ranges use half-open intervals.

Sequence	Source frames	Eval. states	Metric samples
s01_box_use_01	889	1184	1183
s07_box_grab_01	725	965	964
s01_notebook_use_01	669	890	889
s01_laptop_use_01	653	869	868
s01_waffleiron_use_01	605	805	804

Appendix

TABLE VIII: Complete sequences used in the CHORD comparison.

Explore similar work

A Minimalist Retargeting-Guided Reinforcement Learning Recipe for Dexterous Manipulation
Jul 13, 2026Yunhai Feng, Natalie Leung, Jiaxuan Wang +3Robotic Manipulation Policies Cross-Embodiment Retargeting
Cornell University · Amazon FAR
Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration
Jun 22, 2026Xinghao Zhu, Zixi Liu, Shalin Jain +17Contact-Rich Dexterous Manipulation Dexterous Manipulation
NVIDIA
Weave: Learning Whole-Body Dexterous Loco-Manipulation from Human-Object Interactions
Sep 15, 2026Liu Cao, Xingze Wu, Jingzhi Cui +4Contact-Rich Dexterous Manipulation Humanoid Loco-Manipulation
1Tsinghua IIIS · 3Dalian University of Technology · 4The Chinese University of Hong Kong +1

Jul 13, 2026, cs.ROJ/K move · Enter open · S save

A Minimalist Retargeting-Guided Reinforcement Learning Recipe for Dexterous Manipulation

Yunhai Feng, Natalie Leung, Jiaxuan Wang, Lujie Yang+2

Cornell University · Amazon FAR

Robotic Manipulation Policies Cross-Embodiment Retargeting Dexterous Manipulation Whole-Body Control Offline Reinforcement Learning

Open paper AskPDF arXiv

Recent work in humanoid whole-body control has found success with a simple recipe: retarget human motion to robot kinematic references, then train policies via reinforcement learning (RL) to track them. But how does this recipe transfer to dexterous manipulation? The answer is not obvious, as manipulation involves complex, contact-rich dynamics and requires delicate regulation of contact modes and forces. We present REGRIND, a minimalist retargeting-guided RL pipeline that learns dexterous manipulation policies from a single human demonstration. REGRIND retargets human hand-object motion to a robot reference that preserves hand-object spatial and contact relationships, trains a residual RL policy in simulation to track object-centric keypoints along that reference, and transfers the resulting policy zero-shot to hardware with careful system identification. The resulting policies produce fluid, human-like behavior on two different multi-fingered hands across contact-rich tool-use tasks, including operating a pair of scissors and turning a screwdriver. Through systematic hardware experiments, we identify and analyze the key factors that govern sim-to-real transfer in dexterous manipulation, offering practical guidance for retargeting-based learning in contact-rich settings. Videos and code are available at https://yunhaifeng.com/REGRIND.

Abstract

Explore similar work

A Minimalist Retargeting-Guided Reinforcement Learning Recipe for Dexterous Manipulation

Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration

Weave: Learning Whole-Body Dexterous Loco-Manipulation from Human-Object Interactions