[Paper Notes] ReForce: Learning Force-aware Retargeting for Dexterous Manipulation
Published:
This post supports English / 中文 switching via the site language toggle in the top navigation.
TL;DR
ReForce treats human-to-robot retargeting as a contact-reproduction problem. A conventional retargeter supplies a desired robot-hand pose, while human tactile measurements or a learned reference policy supply a desired force for each fingertip. ReForce compares these references with the robot’s current joint state and measured fingertip forces, then predicts a bounded joint update that repairs missing, excessive, or lingering contact.
The method serves as a lightweight closed-loop execution layer; a separate system supplies the visuomotor task reference. Its general force tracker is behavior-cloned from randomized and augmented simulation interactions, then composed with either live teleoperation or an ACT-style motion-and-force reference policy. Deployment on a new object does not require a per-object digital twin or trajectory optimization.
On real paper-cup grasps, ReForce lowers force-tracking error from 0.309 N to 0.247 N for a side grasp and from 0.736 N to 0.474 N for a top grasp relative to direct replay. With learned references, paper-cup force-safe success reaches 70%, compared with 30% for the reference policy and 60% with admittance control. The tongs result is more mixed: ReForce increases contact engagement, while admittance control achieves the best force-safe success, 90% versus 54.5%. The evidence supports ReForce as a promising contact-correction layer, with task-dependent advantages over a tuned analytical controller.
Paper Info
“ReForce: Learning Force-aware Retargeting for Dexterous Manipulation” is by Yuhang Wu, Lingqi Zeng, Changwei Jing, Jianglong Ye, and Xiaolong Wang from UC San Diego. This note covers arXiv:2608.15560v1, posted on August 16, 2026. The project page contains the system overview and real-robot demonstrations. The current version is an arXiv preprint and does not list a conference venue.
1. Retargeting the Interaction
Kinematic retargeting maps human-hand keypoints or joint configurations onto a robot hand. That mapping captures motion intent, yet the physical outcome remains underdetermined. A pose can miss an object because the robot fingers are shorter; it can create too much force because the hardware is stiffer; and a visually similar grasp can slip because contact occurs on a different surface.
This embodiment gap becomes especially visible in force-sensitive manipulation. A paper cup needs enough multi-finger contact to stay secure and enough compliance to avoid deformation. Tongs need contact to appear at the right time and release cleanly. Joint-space similarity alone does not describe these requirements.
Several prior force-aware retargeting methods optimize a robot trajectory inside a simulation model of the demonstrated hand, object, and contacts. That route can produce physically meaningful motion, but it requires a simulation-ready object and performs optimization separately for recorded trajectories. ReForce learns one feedback controller in advance. At execution time, it uses measured contact to adapt the upstream reference online.
The resulting abstraction is:
human motion + human force or learned motion-force policy
↓
pose reference q* and force reference F*
↓
ReForce + robot joint/force feedback (q, F)
↓
corrected hand command
ReForce therefore changes the retargeting target from geometric correspondence alone to motion-and-contact correspondence.
2. Closed-Loop Force-Aware Retargeting
At control step $t$, the robot provides its observed hand configuration and five fingertip normal forces:
[ s_t=(q_t^{\mathrm{obs}},F_t). ]
An upstream source supplies the desired motion-and-contact state:
[ s_t^\star=(q_t^\star,F_t^\star). ]
ReForce also receives the explicit target error:
[ e_t=(q_t^\star-q_t^{\mathrm{obs}},\;F_t^\star-F_t). ]
The policy predicts a bounded joint-command update:
[ \Delta q_t^{\mathrm{cmd}}=\pi_\theta(s_t,s_t^\star,e_t), ]
which is accumulated and clipped to the hardware limits:
[ q_{t+1}^{\mathrm{cmd}} =\operatorname{clip}(q_t^{\mathrm{cmd}}+\Delta q_t^{\mathrm{cmd}},q_{\min},q_{\max}). ]
This residual interface keeps the upstream kinematic reference as a strong prior. The network learns local corrections: close a finger when target force is present but measured force is absent, reduce closure when contact becomes excessive, and release force when the reference falls.
The policy uses a per-finger MLP. The four non-thumb fingers share a $[128,128]$ trunk and use finger-specific $[64]$ heads; the thumb has a separate trunk and head. Each controlled joint update is limited to $0.088$ rad, approximately $5^\circ$. The design is small enough to serve as a fast feedback component and isolates the thumb’s distinct kinematics.
3. Learning a General Force Tracker from Simulation
ReForce is trained from simulated hand-object trajectories containing joint configurations and fingertip forces:
[ {(q_t,F_t)}_{t=1}^{T}. ]
For each time step, the authors average a future window of up to 16 frames to construct post-hoc motion and force targets:
[ q_t^\star=\frac{1}{K_t}\sum_{k=1}^{K_t}q_{t+k}, \qquad F_t^\star=\frac{1}{K_t}\sum_{k=1}^{K_t}F_{t+k}. ]
These targets describe a nearby state that the simulated rollout successfully reached. The supervised action remains the demonstrated one-step update,
[ \Delta q_t^{\mathrm{demo}}=q_{t+1}-q_t. ]
The distinction matters: the future window gives the controller a motion-and-contact goal, while the one-step label teaches how to move toward it. The behavior-cloning loss averages joints within each finger and then weights the five fingers equally:
[ \mathcal L_{\mathrm{ReForce}} =\mathbb E_{t\sim\mathcal D}\left[ \frac{1}{5}\sum_{i=1}^{5}\frac{1}{D_i} \left|\Delta q_t^{\mathrm{demo},(i)}- \Delta q_t^{\mathrm{cmd},(i)}\right|_2^2 \right]. ]
Equal finger weighting prevents a finger with more joints from dominating the objective.
Recovery-Oriented Data Construction
Nominal interaction data alone may teach tracking near successful trajectories without teaching recovery from the errors seen on hardware. ReForce adds Gaussian joint noise, $0.1$ N force noise, and per-finger force dropout with probability 0.30. It also introduces two structured augmentations:
- Reference-stall augmentation freezes selected pose-reference components for force-active fingers while retaining the demonstrated action. This represents a stale motion reference during ongoing contact.
- Pose-drift augmentation moves low-force fingers toward joint limits and recomputes the corrective label. This creates synthetic recovery examples for displaced fingers.
The final mixture uses 70% Base, 20% reference-stall, and 10% pose-drift samples. Object assets and target grasps originate from Dex1B. This detail sharpens the “no digital twin” claim: ReForce still learns from simulated objects, while deployment does not require a matched digital twin for every new target object.
4. Two Sources of Motion-and-Force References
Online Human Teleoperation
In live teleoperation, a Quest controller supplies wrist pose, a Manus glove captures hand pose, and five fingertip FSR sensors measure human force. The human and robot tactile measurements are calibrated into Newtons. Kinematic retargeting produces $q_t^\star$, the human FSRs provide $F_t^\star$, and the XHand’s fingertip sensors close the robot-side loop.
This configuration lets the operator specify task intent and desired contact while ReForce compensates for the embodiment gap during execution.
Offline Learned References
The paper also trains a separate ACT-style reference policy for each task from approximately 30–40 human demonstrations. Its motion branch predicts a 32-step joint-reference chunk from a nine-step configuration history and task phase. Its force branch predicts the corresponding force chunk from task phase alone:
[ {\hat q_{t+k\mid t}^{\mathrm{ref}}}{k=0}^{C-1} =\pi\eta^q(\mathbf Q_t,\phi_t), \qquad {\hat F_{t+k\mid t}^{\mathrm{ref}}}{k=0}^{C-1} =\pi\eta^F(\phi_t). ]
The model uses separate MSE losses for motion and force,
[ \mathcal L_{\mathrm{ACT}} =\mathcal L_q^{\mathrm{MSE}}+\lambda_F\mathcal L_F^{\mathrm{MSE}}, \qquad \lambda_F=1, ]
and temporally ensembles overlapping chunks during deployment. The predicted force is a phase-conditioned contact plan; it does not use measured robot force. ReForce supplies the missing reactive layer by comparing that plan with online tactile feedback.
The ACT model has no visual input. Its reference can encode a nominal task sequence, while its ability to react to object displacement or visual task progress is limited.
5. Real-Robot Evaluation
The hardware platform is a UFACTORY xArm with a 12-DoF XHand and fingertip tactile sensing. Experiments cover side and top paper-cup grasps plus tongs manipulation. The baselines are direct reference replay and a hand-designed task-space admittance controller that maps force error into fingertip displacement and then into joint correction through a damped Jacobian inverse.
Replayed Paper-Cup References
Each method receives the same recorded source trajectory. Force-tracking error is the mean absolute difference between measured and reference force over time and fingertips.
| Grasp | Direct replay | Admittance | ReForce |
|---|---|---|---|
| Side | $0.309\pm0.006$ N | $0.280\pm0.013$ N | $0.247\pm0.036$ N |
| Top | $0.736\pm0.000$ N | $0.619\pm0.048$ N | $0.474\pm0.049$ N |
The comparison uses five trials per condition. Direct replay cannot respond to contact mismatch. Admittance control improves tracking, although its fixed force-to-motion mapping can retain residual force after the target decreases. ReForce obtains the lowest error for both grasp directions.
Learned Motion-and-Force References
Force-safe success requires task completion, force below the task threshold, and no severe missing-contact failure. A trial counts as severe missing contact when at least three fingers never establish contact.
| Task | Method | Force-safe success | Tracking error | Over-force | Severe missing contact | Active fingers |
|---|---|---|---|---|---|---|
| Paper cup | Reference policy | 30.0% | 0.812 N | 0/10 | 7/10 | 0.17 |
| Paper cup | + Admittance | 60.0% | 0.683 N | 4/10 | 1/10 | 1.26 |
| Paper cup | + ReForce | 70.0% | 0.124 N | 2/10 | 0/10 | 2.61 |
| Tongs | Reference policy | 50.0% | 0.454 N | 0/10 | 5/10 | 0.94 |
| Tongs | + Admittance | 90.0% | 0.419 N | 0/10 | 1/10 | 1.34 |
| Tongs | + ReForce | 54.5% | 0.441 N | 4/11 | 1/11 | 1.62 |
Paper-cup grasping is the clearest positive result: ReForce establishes multi-finger contact reliably, sharply reduces force error, and raises force-safe success. Tongs expose the tradeoff. ReForce activates more fingers and avoids most missing-contact failures, yet four over-force trials reduce its force-safe success. Admittance control wins on that task.
The paper’s results therefore support a precise conclusion: ReForce consistently improves contact engagement, while its advantage in force regulation and task success depends on the interaction.
6. What the Ablations Show
The simulation ablation evaluates three random seeds on 14,605 held-out episodes. Adding the pose reference to current state and target force cuts force-tracking error by roughly 32–35%. Adding explicit pose and force errors improves it further.
| Training distribution | State + force target | + pose reference | + error features |
|---|---|---|---|
| Base | 0.0601 N | 0.0400 N | 0.0388 N |
| + reference stall | 0.0592 N | 0.0401 N | 0.0381 N |
| + pose drift | 0.0609 N | 0.0396 N | 0.0385 N |
| + both augmentations | 0.0590 N | 0.0396 N | 0.0379 N |
The largest gain comes from the input representation: force tracking needs the intended pose because a force error alone does not identify which joint motion should create the missing contact. Explicit error features make that relationship easier for a small MLP to learn. Structured recovery data provides a smaller but consistent gain once the full input is available.
7. Strengths and Limitations
ReForce has a clean abstraction boundary. Reference generation decides what interaction should happen; ReForce decides how the current robot hand should adjust to realize it. This makes the controller composable with live teleoperation, replayed demonstrations, and learned reference policies. The compact per-finger network and bounded residual action are also practical choices for high-rate closed-loop control.
The real-world evidence remains narrow. Paper-cup replay uses five trials per condition, and the learned-reference experiments use roughly ten trials per task and method. The two evaluated task families cannot establish broad object, material, hand-hardware, or morphology generalization. A stronger evaluation would include unseen objects, different compliance regimes, disturbances, sensor drift, and multiple dexterous hands.
The force representation contains one normal-force scalar per fingertip. It omits shear force, slip, torque, contact location, and distributed pressure. These signals become important for sliding, rolling, in-hand rotation, and tool interactions. The tongs over-force failures show that stronger contact engagement can conflict with force safety.
The reference policy also has no vision and predicts force only from task phase. It cannot revise the reference using observed object state. ReForce can correct local contact mismatch, but it cannot recover a globally wrong task plan or recognize that an object has moved somewhere unexpected.
Finally, the generalization claim should be interpreted at the controller level. Training uses simulation assets and Dex1B-derived grasps; the paper demonstrates zero-shot execution on the tested real tasks without building their digital twins. It does not yet show universal force retargeting across arbitrary objects and robot hands.
Takeaway
ReForce’s reusable idea is to insert a learned tactile feedback layer between human-guided references and robot control. A pose target provides geometric intent, a force target provides contact intent, and the robot’s measured fingertip force reveals the current physical mismatch. Simulation behavior cloning then learns the local joint corrections that connect the three.
The paper also illustrates an important systems distinction. Better contact engagement and better task success are related but separate objectives. ReForce is strong at making intended contacts appear; the tongs experiment shows that regulating those contacts safely still depends on task dynamics and training coverage. The most promising future version would combine ReForce’s fast tactile correction with visual, task-aware reference generation and richer tactile observations.