Robotics · arXiv:2610.30117 v1 · submitted 21 Oct 2026 · 30 references
CW-WAM: A Structured Contact-Wrench World Action Model for Force-Sensitive Manipulation
The paper says
“Force-sensitive manipulation requires contact and load information that is not provided explicitly by visual observations. We present CW-WAM, a world action model that jointly predicts video, motion, and contact wrench.”
more of the abstract
Inspired by hybrid motion-force control, it contains a unified interaction module that treats both motion and contact wrench as modeling targets and predicts time-aligned motion-wrench chunks as control references. A pair-causal attention mechanism structures the motion-wrench dependencies, while a long fixed-length wrench memory retains past interaction information. The predicted wrench chunk acts as a force reference for a 20-Hz outer-loop admittance controller within a slow-fast control system that maintains force feedback and smooth motion between model queries. CW-WAM improves contact prediction and increases the insertion success rate from 25% to 70% for square pegs with 1-mm clearance per side and from 40% to 95% for power adapters, compared with WAM without wrench input. It also improves lifting-height accuracy across different loads. Our framework with an admittance controller reduces force-reference RMSE from 4.8 to 2.7 N for peg insertion and from 2.0 to 1.3 N for wiping, respectively, with smaller force fluctuations during successful execution. Project website: https://cw-wam-project.github.io/
from the abstract
Pith says
Machine review · deepseek-v4-flash · 22 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A solid, honest empirical paper: structured wrench prediction plus a predicted-wrench admittance reference buys real gains on tight-clearance insertion; the open questions are benchmarking the closest prior predictors, 20-trial statistics, and unvalidated wrist-F/T ground truth. This line is wrong
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 30 references say.
The paper says, at Sec. V-A, Sec. III-A, Eq. (5), Sec. V-D
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The only wrench signal used anywhere in the paper is the output of a single processing chain: sensor bias removal, calibrated tool-gravity subtraction, TCP-frame transform, and a causal second-order Butterworth low-pass filter with a 10-Hz cutoff (Sec. V-A). This compensated wrench is simultaneously the training target (Sec.
more of objection 1
III-A pairs each motion target with it), the controller feedback in Eq. (5), and the basis of every force-quality metric in Sec. V-D, so prediction, control, and evaluation all live inside the same unvalidated signal chain. No sentence reports a free-space residual check, a known-payload gravity-compensation check, a second-sensor comparison, or a raw-versus-filtered trace at contact onset; the stated limitations (Sec. VI) cover only demonstration-distribution scope and generalization. Because the 10-Hz cutoff equals the Nyquist frequency of the 20-Hz admittance loop, a causal second-order filter can phase-delay and attenuate precisely the contact transients (jamming, seating, release) on which the tasks' physical outcomes depend. The physical success claims (insertion, release, ink removal, rim clearance) are robust to this concern since success is judged against task criteria, but the force-reference RMSE reductions claimed in the abstract (4.8 to 2.7 N peg, 2.0 to 1.3 N wiping) and the contact-prediction metrics in Table II compare references and measurements through the same filter, so a smoothed or misaligned wrench could inflate the apparent gains. Please add a free-space residual wrench statistic, a known-payload compensation check, a raw-versus-filtered comparison across a contact onset with an estimate of the filter's phase lag in the contact band, and a timestamp-alignment verification (for example, cross-correlation between commanded motion and wrench transients), or explicitly re-scope the force-quality claims as agreement with the filtered wrist estimate.
The paper says, at Sec. V-E, Fig. 4, abstract
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The headline success improvements (abstract: 25% to 70% for square peg, 40% to 95% for power adapter, with WAM as the baseline) rest on proportions of 20 trials per condition (5/20 versus 14/20 and 8/20 versus 19/20). The paper reports the raw counts, which is commendable, but it attaches uncertainty only to the open-loop metrics (Sec.
more of objection 2
V-C gives bootstrap CIs), not to the robot success rates. With n = 20, the Clopper-Pearson intervals for 25% and 70% success are approximately [8.7%, 49.0%] and [45.7%, 88.1%], and several pairwise contrasts underlying the Sec. V-E narrative are marginal under Fisher's exact test (for example, 7/20 versus 14/20 for peg insertion, the basis of the claim that longer wrench memory improves both insertion tasks, has a two-sided p of roughly 0.06, while the headline 5/20 versus 14/20 contrast is significant at about p = 0.008). Please report exact confidence intervals or exact tests for the success contrasts that support the abstract and Sec. V-E claims.
The paper says, at Sec. V-E, Fig. 4
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The claim that 'in successful peg trials, CW-WAM reduces position-tracking RMSE from 12.5 to 2.8 mm and force-reference RMSE from 15.2 to 4.8 N relative to WAM + Wrench' compares medians over successful subsets of different sizes and composition (8/20 versus 14/20 trials).
more of objection 3
Which trials succeed differs by method, so this contrast reflects a mixture of control quality and trial selection; the paper's own interpretation of the adapter results (higher force-reference RMSE and contact-force variation for CW-WAM, attributed to 'learned search and firm insertion behavior') shows that success and measured force quality interact in exactly this way. The all-trials bars in Fig. 4 partially mitigate the concern, but the reported 12.5-to-2.8 mm and 15.2-to-4.8 N numbers are subset medians; please either base the headline quality comparisons on all trials or report the per-trial distributions and state the selection criterion explicitly.
6 minor objections
The paper says, at Sec. V-A
Pith says: minor objection 1 of 6
'Aligned with the wrench stream by timestamps' is the only description of the 150-Hz-to-15-Hz temporal alignment; please specify the procedure (nearest-neighbor, interpolation, tolerated jitter), since Sec. III-A pairs each motion target with the wrench at 'the same timestamp' and a systematic offset would corrupt all paired targets.
The paper says, at Sec. V-D
Pith says: minor objection 2 of 6
The demonstration-median force reference is computed from the training demonstrations; please state explicitly that the 22 test episodes used for open-loop evaluation and the trial placements used for the robot tests are disjoint from the demonstrations that define the median, since the Fig. 7 comparisons are only meaningful under that condition.
show 4 more
The paper says, at Sec. V-F
Pith says: minor objection 3 of 6
The spring-scale transport results report medians over ten trials per load per method; please add the IQRs or a small statistical test for the load-dependence claim beyond the peak-height summary.
The paper says, at Sec. III-C
Pith says: minor objection 4 of 6
Please state how the shared noise level sigma is sampled across blocks during training and how the four-step Euler inference relates to that schedule.
The paper says, at Sec. V-A
Pith says: minor objection 5 of 6
The operator's force feedback during teleoperation comes from follower external-joint-torque estimates, while the recorded and learned signal comes from the wrist F/T sensor; a sentence relating the two would help readers interpret what the demonstrations contain.
The paper says, at Throughout
Pith says: minor objection 6 of 6
The compiled text renders the model name with inconsistent spacing ('CW-W AM' in the abstract and several headings); this is a formatting artifact but should be fixed in the camera-ready version.
What would settle it
On the same Franka setup, run the four demonstration trajectories in free space with no contact and with known dead weights while recording the raw 150-Hz wrist F/T stream; compare the bias-removed, gravity-compensated, TCP-transformed, 10-Hz-filtered value against the expected external wrench. Then, on the existing peg/adapter datasets, align raw and filtered wrench at first-contact onsets and quantify attenuation and phase lag. If free-space residual exceeds about 1 N, known-payload estimates are off by more than the claimed 1.3-2.7 N RMSE differences, or filtered first-contact rise lags raw by more than one 20-Hz control tick, the wrench-based targets, admittance references, and force metrics are contaminated and the reported gains need re-derivation on an independently validated wrench signal.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 22 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Here's the short version: CW-WAM is a real contribution, and the reporting is more honest than most papers in this space. It takes the DreamZero-style world action model, adds paired motion-wrench history with a pair-causal attention mask (future wrench attends to its paired action, not conversely) and a 10-second wrench memory, and uses the predicted wrench chunk as the reference for a 20-Hz admittance loop. On their hardware, that combination lifts 1-mm-per-side peg insertion from 5/20 to 14/20 and adapter insertion from 8/20 to 19/20 against WAM without wrench input, the predicted-wrench admittance loop cuts force-reference RMSE from 4.8 to 2.7 N in peg insertion, and CW-WAM tracks load-dependent lift height far closer to the demonstration medians than either baseline. The novelty is genuine: FAWAM and Facet-0, the closest cited action-wrench predictors, don't have the paired-history structure or the admittance-reference use.
Credit where due. The ablations share data, backbone, and schedule with bootstrap CIs on the open-loop metrics; the robot results give raw trial counts and IQRs; and the authors volunteer their own negative results — the no-LTM variant keeps lower position error, and CW-WAM tracks force worse than WAM+Wrench in adapter insertion, explained post hoc as search-and-firm-insertion versus one-shot alignment. The scope limits are stated plainly. The central claim is evaluated on held-out windows and fresh robot trials, not defined into existence.
Soft spots, in proportion. First, the two closest prior action-wrench models are cited but never benchmarked, so the marginal value of the specific structure over plain action-wrench prediction is inferred, not measured. Second, the headline success numbers sit on 20 trials per condition with no CIs or significance tests; some comparisons (7/20 vs 8/20 peg insertion) are sampling noise. Third, the stress-test note is right on the facts: every wrench target, reference, and metric comes from the same bias-removed, gravity-compensated, 10-Hz-filtered wrist F/T stream, with no independent ground-truth check. That could flatter the force-RMSE numbers. But it doesn't sink the success-rate claim: those outcomes are operator-verified from video, so the gain is physical; what the unvalidated signal limits is attribution to the wrench channel and the quantitative force claims.
Bottom line: this paper deserves a serious referee. The right ask is a conditional accept — benchmark FAWAM and Facet-0, add trial-level stats, validate the wrench signal against something independent, and release code. I'd bring it to reading group; anyone working on world models or learned force feedback should read it.
The strongest objection
Sec. V-A defines the only wrench signal used anywhere in the paper: sensor bias is removed, calibrated tool gravity is subtracted, the result is transformed to the TCP frame, and a causal second-order Butterworth filter with a 10-Hz cutoff is applied, and this filtered signal "represents external contact and grasped-payload loads." Training pairs every motion target with this compensated wrench at the same timestamp (Sec. III-A); Eq. (5) feeds the measured wrench into the admittance controller with the predicted wrench as reference; and all force-quality metrics in Sec. V-D compare predicted or demonstration references against this same F_meas. The authors' stated limitations [b228], [b230] cover demonstration-distribution scope and generalization, but no sentence in the paper reports validation of the compensated wrench against independent ground truth: no free-space residual check, no known-payload check, no second load cell, and no comparison of filtered vs raw wrench at contact onset. The causal 10-Hz filter cutoff equals the Nyquist frequency of the 20-Hz admittance loop, so it can attenuate and phase-delay exactly the contact transients (jamming, seating, release) that determine task success. If gravity compensation drifts during fast motions, or if the 150-Hz wrench stream is misaligned with the 15-Hz motion/state stream, then targets, references, and evaluation metrics are corrupted by the same artifact, and the reported 25-to-70% and 40-to-95% success gains and 4.8-to-2.7 N and 2.0-to-1.3 N RMSE reductions could reflect predicting a measurement artifact rather than physical contact. Success rates verified from video would remain physical, but their attribution to the wrench channel would be unsupported without an independently validated wrench signal.
The reader's weakest assumption is exactly the link I would stress-test: the wrist F/T signal after bias removal, gravity compensation, and 10-Hz filtering is treated as ground truth for contact and payload. I agree this is the most load-bearing point. I do not think the other concerns raised in the reader's rationale should move the verdict further. The missing FAWAM/Facet-0 benchmarks bound novelty and marginal value, but do not threaten the internal correctness of the WAM-vs-CW-WAM comparison. The success-rate statistics are small, but the raw counts are disclosed and the main peg and adapter gaps are large (5/20 vs 14/20 and 8/20 vs 19/20), so the headline directional claim is not likely to be pure sampling noise. The adapter force-tracking reversal is disclosed in the text with a plausible task-level explanation and does not contradict the central force-RMSE claims, which are for peg insertion and wiping. The wrench-fidelity concern is different: if the filtered, compensated wrench does not faithfully represent the physical interaction, then the model's training targets, the admittance reference in Eq. (5), and every force metric in Fig. 4 are all computed from the same possibly corrupted signal, so the central demonstration of a contact-wrench benefit is not yet established. The authors do not state this as a limitation, so it remains an unresolved condition rather than an acknowledged scope restriction. The proposed test would settle it; if the calibration residuals and filter lag are small, the central claim stands and the conditional verdict is the right bar.
The referee's report
The paper is timely and the engineering is careful: all ablations share data, backbone, and schedule (Sec. V-B); open-loop comparisons carry task-stratified bootstrap CIs (Sec. V-C); robot results report raw trial counts, medians, and IQRs; and the authors honestly report findings that cut against their own narrative, including the no-LTM variant's lower position error and WAM+Wrench's lower adapter force error. The design choice to use the predicted wrench directly as an admittance reference, rather than training a separate corrective policy, is clean and well motivated by the hybrid motion-force control analogy, and the slow-fast execution system is described in enough detail to be reconstructable. Open-loop evaluation on 373 windows from 22 held-out episodes, with bootstrap CIs, addresses the main circularity risk, and the robot success criteria are physical outcomes judged from video rather than quantities defined by the model. If the force-quality claims survive closer scrutiny of the wrench measurement chain, this is a solid contribution to force-sensitive manipulation with world action models.
Pith's reading of the paper (a machine's summary, not the authors')
This paper argues that a robot policy learns contact-rich tasks more reliably when the model is asked to predict not only what it will see and how it will move, but also the contact wrench expected at each motion step. The authors present CW-WAM, a world action model whose interaction module pairs every future motion command with a future wrench prediction, links them with pair-causal attention, and keeps a ten-second memory of past wrench readings. On two 7-DoF robot arms, they report square-peg insertion success rising from 25% to 70% and power-adapter insertion from 40% to 95%, compared with the same model without wrench input. They further report that feeding the predicted wrench as the reference to a 20-Hz admittance controller reduces force-tracking error from 4.8 to 2.7 N in peg insertion and from 2.0 to 1.3 N in wiping. If correct, the insight is that treating force as a first-class prediction target, rather than a side sensor channel, is what makes a world action model contact-aware.
What carries the argument. The load-bearing mechanism is the paired motion-wrench token stream inside a joint denoising transformer. Recent history is encoded as alternating action-wrench tokens (24 motion commands, each paired with the wrench interval around it), and a long-term wrench memory adds 150 wrench-only intervals covering ten seconds. Pair-causal attention uses a lower-triangular mask over the future pairs $[A^{+}_{1}, W^{+}_{1}, \ldots, A^{+}_{H}, W^{+}_{H}]$ so that each wrench attends to its paired action and to earlier pairs, but a future action does not attend to the wrench predicted alongside it. A temporal wrench encoder compresses up to 16 high-rate (150-Hz) samples per motion interval, so contact changes within a 1.6-second action chunk are not lost. At execution, the predicted wrench becomes the reference $\hat{w}$ in a task-weighted admittance law $M\Delta\ddot{x} + C\Delta\dot{x} + K_a\Delta x = -S(e_w + k_I \eta)$ evaluated at 20 Hz against the measured wrench expressed in the nominal TCP frame, with integral accumulation, saturation limits, and C2-quintic bridging that keeps the 1-kHz joint-impedance trajectory continuous.
Core claim. The paper's central claim is that structured co-prediction of motion and contact wrench is what a world action model needs for force-sensitive manipulation. CW-WAM extends a DreamZero-style video-action backbone so that at each observation time it generates a chunk of future actions and a time-aligned chunk of six-axis wrenches, trained jointly with separate prediction heads under flow matching. The paired structure is enforced by pair-causal attention, a lower-triangular mask in which each future wrench attends to its paired action and to all earlier pairs but not conversely, encoding the physical asymmetry that commanded motion produces contact force. A long-term wrench memory of about ten seconds retains payload tension and sustained resistance that a short window loses. At execution, the predicted wrench chunk is not discarded: it becomes the force reference for a 20-Hz outer-loop admittance controller, while nominal motion is tracked at 1 kHz with continuous bridging between model queries. The reported consequence is that insertion and release success rise across two tight-clearance tasks, lifting height adapts to visually identical payloads of different weight, and measured force tracks the intended contact profile more closely.
The premise it rests on. Every training target, predicted reference, and force-quality metric rests on the wrist force-torque reading after bias removal, tool-gravity compensation, and a causal 10-Hz low-pass filter, and the paper does not validate that signal against independent ground truth, so if compensation drifts during fast motion or the filter erases contact transients, the reported force gains could partly reflect predicting a measurement artifact rather than physical contact.
If the paper is right (Pith's inferences, not the authors')
Using the predicted wrench as a control reference removes the need for a separately trained corrective force policy; a standard admittance law closes the force loop between model queries.
Longer wrench memory is a direct contributor to task success: adding the ten-second LTM raises peg insertion from 7/20 to 14/20 trials and adapter insertion from 16/20 to 19/20 in the paper's evaluation.
Better contact prediction transfers to execution: the lower contact-position and orientation errors from pair-causal attention co-occur with higher insertion and release rates.
Wrench prediction also captures load: peak lifting heights separate correctly for visually identical 30 g and 830 g payloads only in the models that receive wrench information.
Even where success saturates, as in 20/20 whiteboard wiping, force accuracy improves, showing that wrench prediction yields a measurable quality axis independent of task completion.
A natural stress test the paper leaves implicit is out-of-distribution contact geometry: running the predicted-wrench admittance loop on unseen chamfers or hole shapes would show whether a wrong predicted reference degrades gracefully or actively misleads the admittance correction.
The pair-causal asymmetry is a reusable domain prior: future action should not be conditioned on the wrench predicted for it, because the robot has not felt it yet; the same masking idea could transfer to tactile or joint-torque world models.
Because every force metric shares the same filtered wrist-wrench stream, the cleanest way to separate prediction skill from measurement skill is to instrument the task with an independent force plate and check whether the reported 4.8-to-2.7 N and 2.0-to-1.3 N reductions survive against true contact force.
The paper's own future-work agenda, learning the feedback-weighting matrix and contact frame alongside the wrench, suggests a testable version of the whole scheme in which the admittance gains themselves become model outputs conditioned on the task.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Force-sensitive manipulation requires contact and load information that is not provided explicitly by visual observations. We present CW-WAM, a world action model that jointly predicts video, motion, and contact wrench. Inspired by hybrid motion-force control, it contains a unified interaction module that treats both motion and contact wrench as modeling targets and predicts time-aligned motion-wrench chunks as control references. A pair-causal attention mechanism structures the motion-wrench dependencies, while a long fixed-length wrench memory retains past interaction information. The predicted wrench chunk acts as a force reference for a 20-Hz outer-loop admittance controller within a slow-fast control system that maintains force feedback and smooth motion between model queries. CW-WAM improves contact prediction and increases the insertion success rate from 25% to 70% for square pegs with 1-mm clearance per side and from 40% to 95% for power adapters, compared with WAM without wrench input. It also improves lifting-height accuracy across different loads. Our framework with an admittance controller reduces force-reference RMSE from 4.8 to 2.7 N for peg insertion and from 2.0 to 1.3 N for wiping, respectively, with smaller force fluctuations during successful execution. Project website: https://cw-wam-project.github.io/
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-22
The authors' main claim, as the classifier picked it out 0.98
“CW-WAM improves contact prediction and increases the insertion success rate from 25% to 70% for square pegs with 1-mm clearance per side and from 40% to 95% for power adapters, compared with WAM without wrench input.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.09 · human participants 0.05 · controlled comparison 0.96 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.97 · mentions code or data availability 0.31
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Force-sensitive manipulation requires contact and load information that is not provided explicitly by visual observations. | background 0.63 | 0.07 | 0.03 | asserted 1.00 |
| We present CW-WAM, a world action model that jointly predicts video, motion, and contact wrench. | contribution 0.99 | 0.92 | 0.08 | asserted 1.00 |
| Inspired by hybrid motion-force control, it contains a unified interaction module that treats both motion and contact wrench as modeling targets and predicts time-aligned motion-wrench chunks as control references. | method 0.97 | 0.81 | 0.06 | asserted 1.00 |
| A pair-causal attention mechanism structures the motion-wrench dependencies, while a long fixed-length wrench memory retains past interaction information. | method 0.99 | 0.82 | 0.03 | asserted 1.00 |
| The predicted wrench chunk acts as a force reference for a 20-Hz outer-loop admittance controller within a slow-fast control system that maintains force feedback and smooth motion between model queries. | method 1.00 | 0.80 | 0.05 | asserted 1.00 |
| CW-WAM improves contact prediction and increases the insertion success rate from 25% to 70% for square pegs with 1-mm clearance per side and from 40% to 95% for power adapters, compared with WAM without wrench input. | result 1.00 | 0.95 | 0.50 | asserted 1.00 |
| It also improves lifting-height accuracy across different loads. | result 1.00 | 0.90 | 0.10 | asserted 1.00 |
| Our framework with an admittance controller reduces force-reference RMSE from 4.8 to 2.7 N for peg insertion and from 2.0 to 1.3 N for wiping, respectively, with smaller force fluctuations during successful execution. | result 1.00 | 0.95 | 0.16 | asserted 1.00 |
| Project website: https://cw-wam-project.github.io/ | future work 0.52 | 0.13 | 0.01 | unclear 0.95 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-22: 1 request, 6,917 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (7)
Showing 6 of the paper's 7 figures. Figures are the authors', reproduced from the paper.
References (30)
- [1] S. Ye, Y . Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y . L. Tan, C. Zhu, J. Xiang,et al., “World action models are zero-shot policies,” arXiv preprint arXiv:2602.15922, 2026. 0 20 Force [N] RMSE 3.88 N (a) No Wrench Tracking 0.0 0.2 T orque [N m] 0 20 Force [N] RMSE... on PithPith's reviewarXiv
- [2] Z. He, H. Fang, J. Chen, H.-S. Fang, and C. Lu, “FoAR: Force- aware reactive policy for contact-rich robotic manipulation,”IEEE Robotics and Automation Letters, vol. 10, no. 6, pp. 5625–5632, 2025. arXiv:2411.15753 on PithPith's reviewarXiv
- [3] G. Lee, Y . Lee, K. Kim, S. Lee, S. Noh, S. Back, and K. Lee, “Manip- Force: Force-guided policy learning with frequency-aware representation for contact-rich manipulation,”arXiv preprint arXiv:2509.19047, 2025 on PitharXiv
- [4] H. Xue, J. Ren, W. Chen, G. Zhang, Y . Fang, G. Gu, H. Xu, and C. Lu, “Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation,” inProceedings of Robotics: Science and Systems, 2025. arXiv:2503.02881 on PithPith's reviewarXiv
- [5] H. He, Z. Yan, Q. Liu, N. Guo, and W. Lian, “FAWAM: Force-aware world action models for closed-loop contact-rich manipulation,”arXiv preprint arXiv:2606.08555, 2026 on PithPith's reviewarXiv
- [6] H. Deng, H. Liu, W. Guo, Y . Ling, Z. Yang, Y . Xue, H. Sun, L. Wang, and Z. Wang, “Facet-0: A robotic foundation model for contact-rich precise manipulation,”arXiv preprint arXiv:2609.01596, 2026 on PithPith's reviewarXiv
- [7] C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. C. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” in Robotics: Science and Systems, 2023 on Pith
- [8] T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” inRobotics: Science and Systems, 2023 on Pith
- [9] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter,et al., “ π0: A vision–language–action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024 on PithPith's reviewarXiv
- [10] K. Blacket al., “ π0.5: A vision–language–action model with open- world generalization,” inConference on Robot Learning, vol. 305 of Proceedings of Machine Learning Research, pp. 17–40, 2025 on Pith
- [11] P. Wu, A. Escontrela, D. Hafner, P. Abbeel, and K. Goldberg, “Day- dreamer: World models for physical robot learning,” inProceedings of the 6th Conference on Robot Learning, vol. 205 ofProceedings of Machine Learning Research, pp. 2226–2240, 2023 on Pith
- [12] Y . Du, S. Yang, B. Dai, H. Dai, O. Nachum, J. Tenenbaum, D. Schuur- mans, and P. Abbeel, “Learning universal policies via text-guided video generation,” inAdvances in Neural Information Processing Systems, vol. 36, pp. 9156–9172, 2023 on Pith
- [13] S. Zhou, Y . Du, J. Chen, Y . Li, D.-Y . Yeung, and C. Gan, “Robodreamer: Learning compositional world models for robot imagination,” inProceed- ings of the 41st International Conference on Machine Learning, vol. 235 ofProceedings of Machine Learning Research, pp. 61885–61896, 2024 on Pith
- [14] P.-C. Ko, J. Mao, Y . Du, S.-H. Sun, and J. B. Tenenbaum, “Learning to act from actionless videos through dense correspondences,”arXiv preprint arXiv:2310.08576, 2023 on PithPith's reviewarXiv
- [15] M. J. Kim, Y . Gao, T.-Y . Lin, Y .-C. Lin, Y . Ge, G. Lam, P. Liang, S. Song, M.-Y . Liu, C. Finn, and J. Gu, “Cosmos Policy: Fine-tuning video models for visuomotor control and planning,”arXiv preprint arXiv:2601.16163, 2026 on PithPith's reviewarXiv
- [16] T. Yuan, Z. Dong, Y . Liu, and H. Zhao, “Fast-WAM: Do world action models need test-time future imagination?,”arXiv preprint arXiv:2603.16666, 2026 on PithPith's reviewarXiv
- [17] Y . Lou, Y . Ye, Y . Fu, J. Cen, X. Chi, Y . Lyu, P. Jia, S. Han, Z. Lu, and S. Zhang, “Dream-tac: A unified tactile world action model for contact- rich robot manipulation,”arXiv preprint arXiv:2606.08737, 2026 on PithPith's reviewarXiv
- [18] S. Tian, Y . Zheng, Y . Zheng, S. Gu, Y . Zang, Y . Qin, W. Li, H. Li, W. Ding, and D. Zhao, “VT-WAM: Visual-tactile world action model for contact-rich manipulation,”arXiv preprint arXiv:2607.02503, 2026 on PithPith's reviewarXiv
- [19] H. Yuan, W. Yi, Z. Zhang, W. Chen, Y . Mo, J. Yin, X. Li, X. Zeng, C. Wen, C. Lu, K. Driggs-Campbell, and I. Lourentzou, “VTAM: Video- tactile-action models for complex physical interaction beyond VLAs,” arXiv preprint arXiv:2603.23481, 2026 on PithPith's reviewarXiv
- [20] J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y . Song, P. Cai, C. Lu, and W. Zhang, “ForceVLA: Enhancing VLA models with a force-aware MoE for contact-rich manipulation,”arXiv preprint arXiv:2505.22159, 2025 on PitharXiv
show 10 more
- [21] W. Xie, S. Caldararu, and N. Correll, “Just add force for contact-rich robot policies,”arXiv preprint arXiv:2410.13124, 2024 on PithPith's reviewarXiv
- [22] M. Wang, Z. Yue, R. Lu, Y . Li, Z. Wang, G. Pan, K. Dong, J. Cheng, Y . Cheng, and H. Liu, “PhaForce: Phase-scheduled visual–force pol- icy learning with slow planning and fast correction for contact-rich manipulation,”arXiv preprint arXiv:2603.08342, 2026 on PitharXiv
- [23] Y . Zheng, S. Gu, Y . Zheng, W. Li, Y . Zang, S. Tian, X. Li, C. Hao, C. Gao, S. Liu, H. Li, Y . Chen, S. Yan, and W. Ding, “OmniVTA: Visuo-tactile world modeling for contact-rich robotic manipulation,”arXiv preprint arXiv:2603.19201, 2026 on PithPith's reviewarXiv
- [24] C. C. Beltran-Hernandez, D. Petit, I. G. Ramirez-Alpizar, T. Nishi, S. Kikuchi, T. Matsubara, and K. Harada, “Learning force control for contact-rich manipulation tasks with rigid position-controlled robots,” IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 5709–5716, 2020 on Pith
- [25] Y . Hou, Z. Liu, C. Chi, E. Cousineau, N. Kuppuswamy, S. Feng, B. Burchfiel, and S. Song, “Adaptive compliance policy: Learning approximate compliance for diffusion guided control,” in2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 4829– 4836, 2025 on Pith
- [26] H. Fang, S. Tang, M. Mei, H. Qin, Z. He, J. Chen, Y . Feng, C. Wang, W. Liu, Z. He, C. Lu, and S. Wang, “Force policy: Learning hybrid force–position control policy under interaction frame for contact-rich manipulation,”arXiv preprint arXiv:2602.22088, 2026 on PithPith's reviewarXiv
- [27] Y . Zhou, C. Barnes, J. Lu, J. Yang, and H. Li, “On the continuity of rotation representations in neural networks,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 5745–5753, 2019 on Pith
- [28] Team Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie,et al., “Wan: Open and advanced large-scale video generative models,”arXiv preprint arXiv:2503.20314, 2025 on PithPith's reviewarXiv
- [29] K. M. Lynch and F. C. Park,Modern Robotics: Mechanics, Planning, and Control. Cambridge University Press, 2017 on Pith
- [30] N. Hogan, “Impedance control: An approach to manipulation: Part I—Theory,”Journal of Dynamic Systems, Measurement, and Control, vol. 107, no. 1, pp. 1–7, 1985 on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment