ReForce: Force-Aware Dexterous Retargeting

• ReForce is a force-aware human-to-robot retargeting method that combines kinematic references with bounded, learned per-joint residual corrections for dexterous manipulation.
• The method uses five robot-side fingertip normal-force channels, desired force targets, motion references, and tracking errors to adapt online during teleoperation or offline demonstration translation.
• ReForce reduced force-tracking error to 0.247 N for side grasps and 0.474 N for top grasps in reported paper-cup experiments, while improving multi-finger contact and addressing morphology, compliance, and friction differences.

ReForce is a force-aware human-to-robot retargeting method for dexterous manipulation. It augments kinematic retargeting with a learned residual policy that uses robot-side fingertip-force feedback to reproduce demonstrated contact behavior despite differences in morphology, actuation, compliance, friction, and grasp geometry. The method predicts bounded joint-command updates rather than direct force, torque, Cartesian, or velocity commands, and supports both online force-aware teleoperation and offline translation of human demonstrations. Its principal formulation and experiments are reported in “ReForce: Learning Force-aware Retargeting for Dexterous Manipulation” (Wu et al., 16 Aug 2026). The name is not unique in the literature: it also denotes a malware-evasion method (Quertier et al., 2022), an adaptive semantic objective for LLM jailbreak auditing (Geisler et al., 24 Feb 2025), and is used in related works on Text-to-SQL, physical control, efficient VLA fine-tuning, and force-budgeted assembly (Deng et al., 2 Feb 2025, Zhang et al., 13 Sep 2025, Zabounidis et al., 3 Nov 2025, Lin et al., 30 Oct 2025, Cavallazzi et al., 4 Jun 2026, Zhang et al., 24 Jun 2026, Rah et al., 23 Jul 2026).

1. Problem and conceptual basis

Human-to-robot retargeting traditionally emphasizes kinematic correspondence. Human joint configurations, fingertip locations, or wrist poses are mapped to robot joint commands, often by minimizing a kinematic discrepancy. Such mappings can reproduce gross hand motion while failing to reproduce the physical interaction demonstrated by the human.

Contact-rich manipulation depends on forces as well as pose. Identical robot trajectories can lead to different outcomes depending on object geometry and material, robot–object and fingertip–object contact locations, finger length and joint arrangement, compliance, actuator behavior, friction, grasp geometry, contact timing, and whether all intended fingers engage. A robot may follow a retargeted trajectory while failing to grasp a paper cup, deforming it through excessive force, or omitting contact with several fingers during tong manipulation.

ReForce therefore uses a layered structure:

$$\text{human motion and force demonstration} \longrightarrow \text{kinematic robot reference} \longrightarrow \text{force-aware residual correction} \longrightarrow \text{robot joint command}.$$

The kinematic stage supplies the intended motion, while the force-aware stage adapts execution according to the discrepancy between desired and measured fingertip forces. The method is trained offline on simulated hand–object interactions but operates online using robot observations. It is intended to reduce the embodiment gap without requiring a task-specific digital twin during deployment.

ReForce is not principally a new reinforcement-learning optimizer. The method is trained through behavior cloning on simulated interaction data, while its force-aware behavior arises from conditioning on desired forces, measured forces, motion references, and force-tracking errors. The paper describes the approach as a learned force tracker and residual retargeting layer.

2. State, force representation, and action formulation

The robot has $D_q$ controllable joints and observes a joint configuration and fingertip-force vector:

$$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$

The force vector contains five scalar normal-force channels corresponding to the thumb, index, middle, ring, and pinky:

$$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$

ReForce currently models fingertip normal force only. It does not use shear force, slip, full contact wrench, or spatial tactile distributions.

The upstream reference consists of a desired robot configuration and desired per-finger force vector:

$s_t^\star=(q_t^\star,F_t^\star).$

The complete policy input is

$$s_t = \left( q_t^{\mathrm{obs}}, F_t, q_t^\star, F_t^\star, q_t^\star-q_t^{\mathrm{obs}}, F_t^\star-F_t \right).$$

All inputs are standardized. Force preprocessing accounts for the concentration of force values near zero and their positive long tail:

$\bar q=\frac{q-\mu_q}{\sigma_q},$

$$\bar F= \frac{ \log\left(1+\max(F,0)/s_F\right)-\mu_F }{ \sigma_F }, \qquad s_F=1.0\ \mathrm{N}.$$

The ReForce policy predicts a bounded joint-command residual:

$\Delta q_t^{\mathrm{cmd}}=\pi_\theta(s_t).$

The residual is accumulated and clipped by hardware joint limits:

$$q_{t+1}^{\mathrm{cmd}} = \operatorname{clip} \left( q_t^{\mathrm{cmd}}+\Delta q_t^{\mathrm{cmd}}, q_{\min},q_{\max} \right).$$

The reported per-joint residual limit is approximately

$D_q$0

Thus, ReForce modifies the kinematic trajectory incrementally rather than replacing it with an unconstrained action. The output is explicitly a robot joint-command update; it is not a Cartesian pose residual, velocity residual, force residual, or torque residual.

The five force channels also organize the policy around finger groups. A finger is considered force-active when

$D_q$1

This structure allows the policy to generate different corrections for different fingers and to respond to missing contact in one finger without requiring identical motion across the hand.

3. Demonstrations and retargeting pipeline

Human demonstrations combine motion and contact information. The reported teleoperation setup uses a Quest controller for wrist pose, a Manus glove for human hand pose, and five fingertip force-sensitive resistor sensors for human contact force.

Human motion is first mapped into robot joint references:

$D_q$2

where $D_q$3 denotes human observations. Human fingertip forces are calibrated in newtons and mapped to corresponding robot fingertips:

$D_q$4

The resulting robot-space reference trajectory is

$D_q$5

The separation between motion and force references is central. Kinematic retargeting expresses the demonstrator’s intended hand motion, whereas force mapping expresses the intended contact intensity. ReForce subsequently adapts the robot action when actual contact differs from the reference.

The method has two operating modes:

1. Online teleoperation: human motion and force are retargeted in real time, robot fingertip forces are measured, and ReForce produces joint-command corrections at each step.
2. Offline data translation: a recorded human trajectory is first retargeted to robot motion and force references; ReForce then processes the trajectory using robot-side force feedback to produce corrected joint commands.

In online operation, the sequence is:

$D_q$6

The policy itself is trained offline, but the correction is closed-loop during execution. ReForce does not optimize the entire trajectory against a task-specific simulator at inference time. Instead, it applies a local, force-conditioned correction layer.

4. Force-tracker training and architecture

Simulated interaction data

ReForce is trained on large-scale simulated hand–object interaction trajectories. The simulations randomize hand configurations, motion trajectories, object contact properties, joint observations, force observations, contact timing, force responses, missing-contact conditions, and residual-force conditions. Object assets and target grasps used in ablations are seeded from Dex1B.

Each simulated trajectory contains joint and force states,

$D_q$7

together with the demonstrated one-step joint update,

$D_q$8

The method constructs future-window references rather than using only instantaneous targets:

$D_q$9

$$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$0

The future horizon is up to 16 frames. These targets provide local information about the direction of the trajectory and the contact state it should approach. They are not action labels; the supervised action remains the one-step demonstrated update.

Behavior-cloning objective

Let $$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$1 be the number of joints in finger group $$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$2. The training loss averages joints within each finger and then weights the five fingers equally:

$$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$3

This prevents fingers with more joints from dominating optimization. Although the policy is trained by behavior cloning, its conditioning on desired force, measured force, and tracking errors causes it to learn force-dependent motion corrections from simulated interactions.

Network structure

The reported policy uses separate processing for the thumb and for the four non-thumb fingers. The four non-thumb fingers share a $$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$4 trunk, while the thumb uses a separate $$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$5 trunk. Finger-specific output heads contain 64 units and use ReLU activations.

Training uses AdamW with learning rate $$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$6, batch size 1024, 20 epochs, and 500 training steps per epoch. The appendix reports a weight-decay value written as $$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$7, but the supplied text identifies this value as typographically ambiguous.

5. Reference generation and robustness augmentations

ReForce can receive references directly from teleoperation or from an ACT-style learned reference policy. The reference policy uses a configuration history

$$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$8

and a task-phase representation

$$q_t^{\mathrm{obs}}\in\mathbb{R}^{D_q}, \qquad F_t\in\mathbb{R}^{D_F}.$$9

The motion branch predicts a future joint-configuration chunk,

$$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$0

while the force branch predicts a corresponding force chunk:

$$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$1

The force branch is conditioned only on phase and therefore represents phase-dependent contact intent learned from demonstrations. Actual force adaptation is delegated to ReForce. The ACT objective is

$$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$2

in the reported configuration. Separate policies are trained for each task using approximately 30–40 human demonstrations. Overlapping action chunks are temporally ensembled, with more recent predictions receiving slightly larger weights.

To expose the tracker to realistic reference and contact failures, training includes structured augmentations:

• Reference-stall augmentation: for 20% of selected samples, parts of the pose reference are held fixed for force-active fingers while the demonstrated action label is retained.
• Pose-drift augmentation: when a finger has low desired force, $$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$3, its observed joints are perturbed toward joint-limit configurations and the affected action label is recomputed.
• Sensor corruption: the base distribution includes Gaussian joint noise, Gaussian force noise with standard deviation $$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$4, and per-finger force dropout with probability $$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$5.

The evaluated mixtures include the base distribution, base plus reference-stall examples, base plus pose-drift examples, and a combined mixture consisting of 70% base data, 20% reference-stall data, and 10% pose-drift data.

These augmentations are intended to train corrections for stale references, inaccurate configurations, missing contact, and residual force discrepancies. They also support transfer from simulation to real hardware by exposing the policy to observation and interaction deviations during training.

6. Evaluation, interpretation, and limitations

Metrics and baselines

The principal force metric is mean absolute force error over time and fingertips:

$$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$6

Real-world evaluation additionally measures force-safe success, over-force trials, severe missing-contact trials, and the mean number of active fingers. The force thresholds are $$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$7 for paper-cup grasping and $$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$8 for tong manipulation. Force-safe success requires task completion, force below the task-specific threshold, and no severe missing-contact failure. A severe missing-contact failure occurs when three or more fingers never establish contact.

The principal baselines are direct replay and hand-designed admittance control. Direct replay executes the nominal retargeted reference without force-dependent correction. Admittance control uses a proportional-derivative force error and converts it into fingertip displacement through a virtual mass–damping–stiffness system, followed by damped Jacobian inverse kinematics. ReForce replaces this fixed force-to-motion mapping with a learned, contact-dependent mapping.

Force-tracking results

On replayed paper-cup references, ReForce produces the lowest reported force-tracking error:

Grasp	Replay	Admittance	ReForce
Side grasp	$$F_t = \left[ F_{t,1},F_{t,2},F_{t,3},F_{t,4},F_{t,5} \right]^\top.$$9	$s_t^\star=(q_t^\star,F_t^\star).$0	$s_t^\star=(q_t^\star,F_t^\star).$1
Top grasp	$s_t^\star=(q_t^\star,F_t^\star).$2	$s_t^\star=(q_t^\star,F_t^\star).$3	$s_t^\star=(q_t^\star,F_t^\star).$4

The reported qualitative behavior is that direct replay can follow motion without reproducing contact, admittance can retain residual force after the desired force decreases, and ReForce more effectively balances contact establishment with force release.

For learned references, ReForce achieves the highest force-safe success for paper-cup grasping, the lowest force-tracking error, no severe missing-contact failures, and the largest mean number of active fingers according to the supplied text. The paper-cup ReForce row in the reported table is corrupted, so its numerical values are not specified. For tongs, admittance achieves higher force-safe success and lower force error, while ReForce produces the largest mean number of active fingers and reduces severe missing-contact failures relative to the reference policy. ReForce nevertheless records four over-force trials in that condition, reducing its force-safe success to 54.5%.

In simulation, the best reported configuration combines base data, reference-stall augmentation, pose-drift augmentation, and the richest error-feature input. It obtains force error

$s_t^\star=(q_t^\star,F_t^\star).$5

over 14,605 held-out episodes across three random seeds. Adding pose references reduces force error by approximately 32–35% relative to state-plus-force-target input, while explicit motion and force errors provide a further improvement.

Interpretation and limitations

ReForce’s principal contribution is to treat retargeting as force-aware closed-loop adaptation rather than geometric pose matching. Human motion supplies movement intent and human fingertip forces supply contact intent. The robot executes a kinematic approximation, while the learned residual uses robot-side force feedback to compensate for embodiment and interaction differences.

The results indicate that ReForce can improve multi-finger contact engagement and force reproduction in paper-cup and tong manipulation. Its design also supports offline data translation and online teleoperation without requiring a task-specific digital twin during deployment.

The method has several limitations:

• Normal-force representation: it uses five scalar fingertip normal-force channels and does not represent shear force, slip, contact distribution, contact-patch geometry, or the full six-dimensional contact wrench.
• Reference-policy observability: the ACT reference policy uses joint-state history and task phase but no visual observations or explicit task-state feedback. It cannot therefore adapt its nominal motion-and-force reference to visual changes in object pose or task progress.
• Implicit contact reasoning: the method does not explicitly solve contact-point matching or reason over object geometry. Multi-finger and contact-point effects are learned implicitly from simulated interaction data.
• Simulation dependence: training requires large-scale randomized simulated interaction data, although deployment does not require a task-specific digital twin.
• Morphological transfer: a correction learned for one robot hand may not transfer directly to substantially different hardware without suitable training data.
• Task and phase dependence: force references are partly phase-conditioned, and the release or contact behavior may be sensitive to task timing and execution phases.
• Evaluation scope: the reported experiments focus on paper-cup grasping and tong manipulation. The evidence does not establish general performance across arbitrary dexterous tasks.
• Computational reporting: exact inference latency, hardware utilization, and control-loop timing are not provided.

ReForce should therefore be understood as a learned force-aware retargeting layer rather than a complete tactile manipulation architecture. Its experimentally supported contribution is narrower but technically specific: bounded, per-joint residual corrections conditioned on desired and measured fingertip forces can improve contact reproduction and multi-finger engagement when conventional kinematic retargeting is insufficient.