Robotics · arXiv:2611.21099 v1 · submitted 16 Nov 2026 · 50 references
ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models
The paper says
“Frontier multimodal models can now control robots zero-shot from images, yet many manipulation tasks depend on contact forces that images cannot show. Bringing force to an API-only frozen model raises two challenges.”
more of the abstract
First, contact evolves much faster than an API model can decide, so the system must close the loop on force beneath the model. Second, a frozen model cannot interpret raw force streams, so force must be returned in a form the model can reason about. Therefore, we present ForceHarness, a harness that lets frozen models use force without demonstrations or finetuning. After calibrating hand force from a wrist sensor or joint torques alone, ForceHarness addresses the first challenge with self-terminating force primitives that regulate force on their own and return at a contact event. To address the second challenge, a physics translation layer turns each force trace into physical quantities with uncertainty as the tool result. To evaluate ForceHarness against expert force controllers and learned policies under identical conditions, we also build HiddenPhys, a seven-task MuJoCo benchmark with hidden physical variables. On a real arm without a wrist force sensor, ForceHarness raises the zero-shot success rate of GPT-6 Astra from 3.3% to 90.0%. In simulation, it adds 17.5 percentage points on the public ForceBench suite and 34.3 percentage points on HiddenPhys, surpassing force-aware policies trained on demonstrations on both benchmarks and matching hand-written expert controllers on HiddenPhys. Ablations further show that smaller models make significantly better decisions when given physical quantities rather than raw force values.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 17 Nov 2026 · read the full text (182,997 characters) · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. ForceHarness gives frozen frontier models a practical way to use force feedback for contact-rich manipulation, with strong empirical results on a real robot and in simulation. This line is wrong
10 minor · 3 open questions
How often objections from reviews like this one are right: not yet measured
The paper, first character to last
The reviewer was given the whole text, 182,997 characters. Not checked: figures, code, what the 50 references say.
10 minor objections
The paper says, at §3.2
Pith says: minor objection 1 of 10
Equation (1) uses the pseudoinverse of J^T. The paper should note that this assumes J is full row rank (i.e., the robot is not at a singularity). The supplementary material mentions that a Jacobian near a singularity ends the episode with a fault, but this should be stated alongside the equation.
The paper says, at §4
Pith says: minor objection 2 of 10
Table 2 lists 'Peg' with hole position error '15 mm error per axis' but the task card in the supplementary material says '15 mm Gaussian error per axis' and later '25 mm' for the shift condition. The table should be consistent with the supplementary material.
show 8 more
The paper says, at §5.1
Pith says: minor objection 3 of 10
The real-robot tasks are described only briefly. The supplementary material (Table I1) gives more detail, but the main text should at least mention that the tasks include wiping, screw driving, peeling, USB insertion, bottle sorting, and soft/hard discrimination. Currently the main text only lists them in Table 3 without description.
The paper says, at §5.3
Pith says: minor objection 4 of 10
The HiddenPhys benchmark uses 'noisy text' instead of images for the frozen models, while learned policies see two camera views. This is a reasonable design choice to isolate force from vision, but it should be stated more prominently in the main text rather than only in the supplementary material.
The paper says, at §6
Pith says: minor objection 5 of 10
The ablation study (Table 6) compares seven feedback forms but only on HiddenPhys. The paper should note that the relative ordering of forms might differ on real hardware or ForceBench, especially given the different noise characteristics.
The paper says, at Supplementary §A.4
Pith says: minor objection 6 of 10
The pseudocode for 'search' (Fig. A.5) uses a variable 'c' that is set to 'p' on surface contact, but the spiral center is defined relative to 'c'. The description should clarify whether the spiral center is fixed at the first contact point or updated during the search.
The paper says, at Supplementary §B.2
Pith says: minor objection 7 of 10
The center-of-mass offset calculation (Eq. B.2) uses a sign constant s_tau = -1. The paper should explain how this constant is determined (e.g., from the simulator's known lever direction) and whether it is the same for all backends.
The paper says, at Supplementary §C.2
Pith says: minor objection 8 of 10
The task card for Lock says 'handle already held' but the main text (Table 2) says the robot must 'Pull below 15 N'. The task card also mentions a pre-grasp. This should be consistent: is the handle grasped before the episode starts or must the robot grasp it?
The paper says, at Supplementary §D.4
Pith says: minor objection 9 of 10
Figure D.2 shows results for eight models but only on 5 seeds per task. The paper should state the statistical significance of the improvements for smaller models, especially since the sample size is small.
The paper says, at Supplementary §H
Pith says: minor objection 10 of 10
The ForceBench port uses a websocket policy and an adapter. The paper should clarify whether the adapter introduces any latency or approximation beyond the primitives themselves.
Open questions (3)
Pith raised these and could not confirm them against the text of the paper it read. They are questions for the authors, not findings.
- §5.1 Real-robot experiments Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.
- §5.2 ForceBench Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.
- §5.3 HiddenPhys / Table 2 Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.
What would settle it
Running the same six real-robot tasks with ForceHarness on an FR3 arm while systematically varying the joint torque calibration (e.g., skipping the tare step, reducing the settling time, or using absolute force values instead of relative changes) would show whether the calibration procedure is necessary for the reported success rates. If success rates remain high without these calibration steps, the claim that the force estimate is the load-bearing component would be weakened.
from Pith's summary of the paper
How the review went
5 passes by deepseek-v4-flash · 17 Nov 2026
- Reader holds up, with conditions
- Skeptic concern not confirmed: listed as an open question
- Referee 10 minor
- Check against the paper kept 0 of 4 objections
- Editor wrote the verdict above
Not run on this paper: the simulated authors' defence.
Pith's passes raised 4 objections. A check against the paper's text kept 0 and left 4 as questions.
The review in full
The review in brief, machine-written
This paper presents ForceHarness, a three-layer system that lets API-only frozen models (GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.2) use wrist force for robot manipulation without any training or demonstrations. The key idea is to close the force loop beneath the model with self-terminating primitives (guarded moves, sliding, insertion, etc.) and to return force information to the model as calibrated physical quantities with uncertainties rather than raw sensor streams. This is a clean, practical design that addresses a real limitation of vision-only zero-shot policies.
The evaluation is thorough: six real-robot tasks on a Franka FR3 without a wrist sensor (force estimated from joint torques), the public ForceBench suite, and a new MuJoCo benchmark (HiddenPhys) with seven tasks. On the real robot, ForceHarness raises GPT-6 Astra from 3.3% to 90.0% zero-shot success. In simulation it matches or exceeds hand-written expert controllers and outperforms trained force-aware policies (Diffusion Policy, π0.5, ForceVLA) that require hundreds of demonstrations. The ablations are informative, showing that self-terminating primitives alone improve safety and that physical quantities help smaller models make better decisions than raw force values.
The soft spots are real but not fatal. The real-robot evaluation uses only 10 episodes per condition, which is small. The ForceBench evaluation also uses 10 episodes per task for zero-shot methods, and the method was tuned on the evaluation seeds. The HiddenPhys benchmark is new and designed by the authors, so there is a risk of overfitting. The comparison to trained baselines on HiddenPhys may not reflect the strongest possible implementations. The paper does not provide code or data at the time of review. These concerns are addressable, and the paper is otherwise sound.
This paper is for researchers working on zero-shot robot control, force-aware manipulation, and harnesses for frozen models. It deserves a serious referee. The method is clearly described, the results are impressive, and the limitations are acknowledged. I recommend acceptance pending release of code and data, and ideally a larger-scale real-robot evaluation.
The stress test's concern
Not confirmed. Pith could not check this concern against the text of the paper it was given; it is listed as an open question, not a finding.
The referee's report
Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 3 of its 3 major objections did not survive that check.
Pith's reading of the paper (a machine's summary, not the authors')
The paper presents ForceHarness, a three-layer system that enables frozen frontier models (like GPT-6 Astra) to use wrist force for contact-rich manipulation without any demonstrations or finetuning. The key insight is to close the force loop beneath the model using self-terminating primitives that regulate force on their own, and to return force information as calibrated physical quantities rather than raw sensor streams. On a real robot arm without a wrist sensor, ForceHarness raises zero-shot success from 3.3% to 90.0% across six tasks. In simulation, it matches hand-written expert controllers and outperforms trained force-aware policies on a new benchmark called HiddenPhys.
What carries the argument. The three-layer harness: (1) force sensing and calibration that estimates hand force from joint torques via least-squares wrench recovery and removes drift by taring at rest, settling before reading, and reporting changes relative to a reference; (2) twelve self-terminating force primitives (e.g., move_until_contact, push, slide, search, insert, lift, pull) that run admittance control on top of the robot's stock impedance controller and stop on contact events rather than on a clock; (3) a physics translation layer that converts each primitive's force trace into physical quantities (mass, friction, contact height, stiffness, etc.) with uncertainty estimates and status flags, using fixed estimators based on averages and least-squares fits.
Core claim. ForceHarness demonstrates that a frozen frontier model can understand and use force for contact-rich manipulation by placing three layers between the model and the robot: force calibration from joint torques, self-terminating force primitives that regulate contact beneath the model, and a physics translation layer that returns physical quantities with uncertainty. This design lets the model call force tools and read results like payload mass, friction, and contact height, while the low-level control loop runs at 20 Hz without waiting for the model. On a real FR3 arm without a wrist sensor, GPT-6 Astra with ForceHarness achieves 90.0% zero-shot success across six tasks (vs 3.3% without force). On the HiddenPhys benchmark, it matches hand-written expert controllers at 135/140 successes and surpasses trained policies like Diffusion Policy and ForceVLA.
The premise it rests on. The force estimate from joint torques on the real arm (FR3 without a wrist sensor) is accurate enough after taring at rest, settling before reading, and reporting changes relative to a reference to support the reported 90.0% success rate across six tasks and 60 episodes.
If the paper is right (Pith's inferences, not the authors')
Frozen frontier models can perform contact-rich manipulation tasks that require force sensing, without any task-specific training or demonstrations, by using a harness that abstracts force into callable primitives and readable physical quantities.
The self-terminating primitive design decouples the slow decision rate of API models (seconds) from the fast force control loop (20 Hz), making it practical to use remote models for contact tasks.
ForceHarness matches or exceeds hand-written expert controllers on the HiddenPhys benchmark, suggesting that the combination of force primitives and physics translation can replace engineered force control for many tasks.
Smaller models benefit significantly from receiving computed physical quantities rather than raw force values, indicating that the physics translation layer is crucial for making force accessible to models that cannot interpret raw sensor streams.
The approach transfers across multiple frontier models (GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.2) and across real hardware and simulation without modification to the model weights.
The self-terminating primitive design could be extended to other sensing modalities (e.g., tactile, proximity) to give frozen models access to a wider range of physical interactions without training.
The physics translation layer's approach of returning quantities with uncertainty could serve as a template for making any sensor stream interpretable by language models, beyond just force.
The success on the real arm without a wrist sensor (using only joint torque estimates) suggests that many existing robots without dedicated force sensors could benefit from this approach, though the calibration procedure's reliance on taring and relative comparisons may limit precision for tasks requiring absolute force measurements.
The fixed menu of primitives and hand-written instructions are current limitations; future work could explore automatic primitive synthesis or instruction generation to reduce human effort in setting up new tasks.
Check this review
What was read
The passes were given all 182,997 characters of the paper's LaTeX source.
Fingerprint of that text (SHA-256): 2cff42eb16af3b314347280a72f12605ed155d68195e554145eeeeef4acb47e1
Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:
PYTHONPATH=. python scripts/review_input_check.py 2611.21099
The passes, in the order they ran
- Reader deepseek-v4-flash · medium effort · $0.0062 · instructions of the Reader, fingerprint f9ead7300382…
- Skeptic deepseek-v4-flash · medium effort · $0.0056 · instructions of the Skeptic, fingerprint 8f88c0f61488…
- Referee deepseek-v4-flash · medium effort · $0.0059 · instructions of the Referee, fingerprint a852d99aea0d…
- Check against the paper deepseek-v4-flash · medium effort · $0.0099 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
- Editor deepseek-v4-flash · medium effort · $0.0056 · instructions of the Editor, fingerprint e7211f6a3d27…
- Circularity check deepseek-v4-flash · medium effort · $0.0066 · instructions of the Circularity check, fingerprint e9946e39d67b…
- Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0074 · instructions of the Summary for the page, fingerprint e4f56bf568e7…
The passes cost $0.047 in model calls together.
Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00
Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.
curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum
prints the fingerprint of the reader's instructions. The same works for every pass.
The receipt
The review's fingerprint was taken on 17 Nov 2026. The review you are reading still matches it.
Fingerprint of the review (SHA-256): 5000b6cd01ff5f7dd0c5a007c5330ac22f985a8985c4c82ccbb4000bd631054c
curl -s 'https://pith.science/api/v1/verdicts/2611.21099?form=canonical' | sha256sum
prints the fingerprint of the review as it stands now. The rule, in four steps.
A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.
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
Frontier multimodal models can now control robots zero-shot from images, yet many manipulation tasks depend on contact forces that images cannot show. Bringing force to an API-only frozen model raises two challenges. First, contact evolves much faster than an API model can decide, so the system must close the loop on force beneath the model. Second, a frozen model cannot interpret raw force streams, so force must be returned in a form the model can reason about. Therefore, we present ForceHarness, a harness that lets frozen models use force without demonstrations or finetuning. After calibrating hand force from a wrist sensor or joint torques alone, ForceHarness addresses the first challenge with self-terminating force primitives that regulate force on their own and return at a contact event. To address the second challenge, a physics translation layer turns each force trace into physical quantities with uncertainty as the tool result. To evaluate ForceHarness against expert force controllers and learned policies under identical conditions, we also build HiddenPhys, a seven-task MuJoCo benchmark with hidden physical variables. On a real arm without a wrist force sensor, ForceHarness raises the zero-shot success rate of GPT-6 Astra from 3.3% to 90.0%. In simulation, it adds 17.5 percentage points on the public ForceBench suite and 34.3 percentage points on HiddenPhys, surpassing force-aware policies trained on demonstrations on both benchmarks and matching hand-written expert controllers on HiddenPhys. Ablations further show that smaller models make significantly better decisions when given physical quantities rather than raw force values.
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 · 2026-11-17
The authors' main claim, as the classifier picked it out 0.80
“In simulation, it adds 17.5 percentage points on the public ForceBench suite and 34.3 percentage points on HiddenPhys, surpassing force-aware policies trained on demonstrations on both benchmarks and matching hand-written expert controllers on HiddenPhys. Ablations further show that smaller models make significantly better decisions when given physical quantities rather than raw force values.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.41 · human participants 0.03 · controlled comparison 0.92 · randomized assignment 0.03 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.06
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 |
|---|---|---|---|---|
| Frontier multimodal models can now control robots zero-shot from images, yet many manipulation tasks depend on contact forces that images cannot show. | background 0.83 | 0.06 | 0.03 | asserted 0.95 |
| Bringing force to an API-only frozen model raises two challenges. | gap 0.56 | 0.13 | 0.02 | asserted 0.99 |
| First, contact evolves much faster than an API model can decide, so the system must close the loop on force beneath the model. | background 0.60 | 0.10 | 0.03 | asserted 1.00 |
| Second, a frozen model cannot interpret raw force streams, so force must be returned in a form the model can reason about. | background 0.58 | 0.09 | 0.03 | asserted 1.00 |
| Therefore, we present ForceHarness, a harness that lets frozen models use force without demonstrations or finetuning. | contribution 1.00 | 0.89 | 0.27 | asserted 1.00 |
| After calibrating hand force from a wrist sensor or joint torques alone, ForceHarness addresses the first challenge with self-terminating force primitives that regulate force on their own and return at a contact event. | method 0.97 | 0.90 | 0.08 | asserted 1.00 |
| To address the second challenge, a physics translation layer turns each force trace into physical quantities with uncertainty as the tool result. | method 0.96 | 0.91 | 0.05 | asserted 1.00 |
| To evaluate ForceHarness against expert force controllers and learned policies under identical conditions, we also build HiddenPhys, a seven-task MuJoCo benchmark with hidden physical variables. | contribution 0.52 | 0.95 | 0.03 | asserted 0.96 |
| On a real arm without a wrist force sensor, ForceHarness raises the zero-shot success rate of GPT-6 Astra from 3.3% to 90.0%. | result 1.00 | 0.94 | 0.30 | asserted 1.00 |
| In simulation, it adds 17.5 percentage points on the public ForceBench suite and 34.3 percentage points on HiddenPhys, surpassing force-aware policies trained on demonstrations on both benchmarks and matching hand-written expert controllers on HiddenPhys. Ablations further show that smaller models make significantly better decisions when given physical quantities rather than raw force values. | result 1.00 | 0.96 | 0.53 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97
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, 2026-11-17: 1 request, 7,679 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 (4)
Figures are the authors', reproduced from the paper.
References (50)
- [1] Christopher G. Atkeson, Chae H. An, and John M. Hollerbach. Estimation of inertial parameters of manipulator loads and links.Int. J. Robot. Res., 5(3):101–119, 1986 on Pith
- [2] Antonio Bicchi, J. Kenneth Salisbury, and David L. Brock. Contact sensing from force measurements.Int. J. Robot. Res., 12(3):249–262, 1993 on Pith
- [3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, et al. π0: A vision-language- action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024 on PithPith's reviewarXiv
- [4] Kaylee Burns, Ajinkya Jain, Keegan Go, Fei Xia, Michael Stark, Stefan Schaal, and Karol Hausman. GenCHiP: Gen- erating robot policy code for high-precision and contact-rich manipulation tasks. InIEEE/RSJ Int. Conf. Intell. Robots Syst., pages 9596–9603, 2024 on Pith
- [5] Yanzhe Chen, Zechen Bai, Zhijun Cao, Wenzheng Zeng, Kevin Qinghong Lin, Yiqi Lin, Guoqiang Liang, Kevin Yuchen Ma, Qiming Huang, and Mike Zheng Shou. Show-Harness: Just a VLM agent can play robots.arXiv preprint arXiv:2609.10522, 2026 on PithPith's reviewarXiv
- [6] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffu- sion policy: Visuomotor policy learning via action diffusion. InRobot.: Sci. Syst., 2023 on Pith
- [7] Zhaoyang Chu, Earl T. Barr, Claire Le Goues, Peter O’Hearn, Mark Harman, Federica Sarro, and He Ye. OpenRUA: Robot- use agents are zero-shot visuomotor policies.arXiv preprint arXiv:2610.02459, 2026 on PithPith's reviewarXiv
- [8] Zijie Diao, Yitong Chen, Sicheng Xie, Tianyi Lu, Wujian Peng, Guojin Zhong, Houze Xu, Ziyi Ye, Zuxuan Wu, and Yu-Gang Jiang. LIBERO-Agent: Evaluating general-purpose agents for direct embodied manipulation.arXiv preprint arXiv:2609.39507, 2026 on PithPith's reviewarXiv
- [9] Letian Fu, Justin Yu, Karim El-Refai, Ethan Kou, Haoru Xue, Huang Huang, et al. CaP-X: A framework for benchmarking and improving coding agents for robot manipulation. InInt. Conf. Mach. Learn., 2026 on Pith
- [10] Liane Galanti, Dhruv Shah, and Tri Dao. Addressing the orchestration gap in generalist robots via physical agency. arXiv preprint arXiv:2607.21725, 2026 on PithPith's reviewarXiv
- [11] Galbot Team, Xuchuan Chen, Xiaoqian Cheng, Yu Deng, Lihe Ding, Shaocong Dong, et al. Systematically exploring the capabilities of GPT-6 Astra as embodied policies.arXiv preprint arXiv:2609.38537, 2026 on PithPith's reviewarXiv
- [12] Bohan Gan, Xuanzhang Wen, Yongsheng Zhao, Baoping Cheng, Wenhe Jia, Ye Wang, et al. TAO-Force: Unifying force-aware perception and fast-slow control for contact-rich manipulation.arXiv preprint arXiv:2609.18497, 2026 on PithPith's reviewarXiv
- [13] Shijia Ge, Alex Zhou, Jianshu Zeng, Yexing Wan, Di Wu, Zelin Zheng, Yazhe Wang, Zhiqi Jia, Xuan Shangguan, Jay Zhu, Yijun Liu, Lingyu He, Sihang Wu, Xiao He, and Hongcheng Gao. Make code as policy great again: Fron- tier agents write, call, and evolve robot tools.arXiv preprin... on PithPith's reviewarXiv
- [14] Xuan Han, Zinuo Jin, Tianlin Li, Yunfan Lou, Xin Liu, Yuchen Xue, Yuanpeng He, Boxian Yao, Shuhan Zhan, Rui- hai Wu, and Zhibo Pang. ForceBench: Benchmarking force integration for robotic manipulation.arXiv preprint arXiv:2610.19604, 2026 on PithPith's reviewarXiv
- [15] Zihao He, Hongjie Fang, Jingjing Chen, Hao-Shu Fang, and Cewu Lu. FoAR: Force-aware reactive policy for contact- rich robotic manipulation.IEEE Robot. Autom. Lett., 10(6): 5625–5632, 2025 on Pith
- [16] Neville Hogan. Impedance control: An approach to manipu- lation: Part I—Theory.J. Dyn. Syst. Meas. Control, 107(1): 1–7, 1985 on Pith
- [17] Kairui Hu, Siyuan Hu, Fangzhou Hong, Zhaoxi Chen, and Ziwei Liu. Embodied Turing machines: Stateful code for robot recursive self-improvement.arXiv preprint arXiv:2610.12369, 2026 on PithPith's reviewarXiv
- [18] Yang Li, Zhaxizhuoma, Hongru Jiang, Junjie Xia, Hongquan Zhang, Jinda Du, et al. ForceVLA2: Unleashing hybrid force-position control with force awareness for contact-rich manipulation. InIEEE Conf. Comput. Vis. Pattern Recog., 2026 on Pith
- [19] Zexi Li, Yehang Zhang, Wenqian Li, Haojian Huang, Chenxu Wang, Shiyuan Deng, et al. Robo-Harness K1: Harnessing robot-use agents via perception augmentation.arXiv preprint arXiv:2609.29389, 2026 on PithPith's reviewarXiv
- [20] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Haus- man, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In IEEE Int. Conf. Robot. Autom., pages 9493–9500, 2023 on Pith
show 30 more
- [21] Fangcheng Liu, Yeqing Shen, Anda Cheng, Weishi Mi, Chao Tang, Chenyuan Liu, Yushun Xiang, Tingguang Li, Yong-Lu Li, and Yehui Tang. RoboICL: Embodied in-context learning with GPT-6 Astra.arXiv preprint arXiv:2609.34261, 2026 on PithPith's reviewarXiv
- [22] Ruaridh Mon-Williams, Gen Li, Ran Long, Wenqian Du, and Christopher G. Lucas. Embodied large language models enable robots to complete complex tasks in unpredictable environments.Nat. Mach. Intell., 7(4):592–601, 2025 on Pith
- [23] Michael Noseworthy, Bingjie Tang, Bowen Wen, Ankur Handa, Chad Kessens, Nicholas Roy, Dieter Fox, Fabio Ramos, Yashraj Narang, and Iretiayo Akinola. FORGE: Force-guided exploration for robust contact-rich manipula- tion under uncertainty.IEEE Robot. Autom. Lett., 10(5): 4436–4... on Pith
- [24] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. π0.5: A vision-language-action model with open-world generalization. InConf. Robot Learn., 2025 on Pith
- [25] Kyupaeck Jeff Rah and Midum Oh. FORGE-plus: Force- budgeted recovery for contact-rich assembly with a frozen LLM supervisor.arXiv preprint arXiv:2607.21227, 2026 on PithPith's reviewarXiv
- [26] Marc H. Raibert and John J. Craig. Hybrid position/force control of manipulators.J. Dyn. Syst. Meas. Control, 103(2): 126–133, 1981 on Pith
- [27] Robocurve. GPT-6 Astra on robot arms. Third-party evalua- tion report, https://openai.robocurve.org/gpt- 6-astra/, 2026 on Pith
- [28] Xiao Tan, Yixuan Wang, Zhexi Wang, and Wenzhen Yuan. FORCA: A force-aware coding agent for contact-rich robot manipulation.arXiv preprint arXiv:2610.31841, 2026. 9 on PithPith's reviewarXiv
- [29] Kazutoshi Tanaka. TacZero: Training-free peg insertion using a general-purpose vision-language model with tactile feed- back.arXiv preprint arXiv:2610.07621, 2026 on PithPith's reviewarXiv
- [30] Wenqing Tian, Zeyu Zhang, Zhaocheng Liu, Fengwei Liu, Qiang Liu, and Liang Wang. PhysEvo: Astra can act, let it. arXiv preprint arXiv:2610.08995, 2026 on PithPith's reviewarXiv
- [31] Chengcheng Wang, Yuheng Shi, Siyu Xu, Yutong Wang, Jin- jing Zhao, Chenghao Xia, et al. Auriga technical report: A task-agnostic agent harness for zero-shot embodied manipu- lation.arXiv preprint arXiv:2611.01112, 2026 on PithPith's reviewarXiv
- [32] Jiayu Wang, Yue Yu, Bin Zhu, Zhiyao Yang, and Jingjing Chen. SpatialHarness: Test-time spatial scaffolding for fine robotic manipulation.arXiv preprint arXiv:2610.12457, 2026 on PithPith's reviewarXiv
- [33] Yi Wang, Wendi Chen, Zimo Wen, Han Xue, Xueqi Li, Wenye Yu, et al. Never too late for force: Accelerating VLA post- training with reactive force injection. InConf. Robot Learn., 2026 on Pith
- [34] Yubang Wang, Zezheng Huai, Yifei Zhang, Jinhao Wu, Yi- tong Chen, Haozhe Zhang, et al. Embodied task agents can perform fine manipulation through tactile in-context learning. arXiv preprint arXiv:2610.27613, 2026 on PithPith's reviewarXiv
- [35] Ziyong Wang, Yupei Lin, Runxuan Huang, Liang Lin, Dongyu Zhang, and Guangrun Wang. GPT-6-Astra under- stands touch and outperform vision–tactile–language–action models.arXiv preprint arXiv:2610.28457, 2026 on PithPith's reviewarXiv
- [36] Zhijie Wei, Ferris Tan, Jinghui Wang, and Hao Zhou. Scale and selection: What makes automatic harness evolution work for visual-interface robot agents.arXiv preprint arXiv:2609.39304, 2026 on PithPith's reviewarXiv
- [37] Sicheng Xie, Yitong Chen, Haidong Cao, Shunlin Lu, Zux- uan Wu, and Yu-Gang Jiang. Explore, execute, evolve: A skill acquisition and reuse loop for embodied agents.arXiv preprint arXiv:2609.37810, 2026 on PithPith's reviewarXiv
- [38] Chen Yang, Linzhe Shi, Hang Zhang, Mu Xu, Jiansheng Fan, and Chen Wang. GPT-6 Astra is a native visuo- tactile policy for contact-rich manipulation.arXiv preprint arXiv:2610.26846, 2026 on PithPith's reviewarXiv
- [39] Jiawen Yu, Hairuo Liu, Qiaojun Yu, Jieji Ren, Ce Hao, Haitong Ding, Guangyu Huang, Guofan Huang, Yan Song, Panpan Cai, Cewu Lu, and Wenqiang Zhang. ForceVLA: Enhancing VLA models with a force-aware MoE for contact- rich manipulation. InAdv. Neural Inform. Process. Syst., 2025 on Pith
- [40] Heng Zhang, Wei-Hsing Huang, Gokhan Solak, and Arash Ajoudani. OmniVIC: A self-improving variable impedance controller with vision-language in-context learning for safe robotic manipulation. InIEEE-RAS Int. Conf. Humanoid Robots, 2026 on Pith
- [41] Heng Zhang, Wei-Hsing Huang, Qiyi Tong, Gokhan Solak, Puze Liu, Kaidi Zhang, Sheng Liu, Jan Peters, Yu She, and Arash Ajoudani. CompliantVLA-adaptor: VLM-guided vari- able impedance action for safe contact-rich manipulation. arXiv preprint arXiv:2601.15541, 2026 on PitharXiv
- [42] Shuoheng Zhang, Yifu Yuan, Hongyao Tang, Yan Zheng, Qiaojun Yu, et al. ForceFlow: Learning to feel and act via contact-driven flow matching.arXiv preprint arXiv:2605.11048, 2026 on PithPith's reviewarXiv
- [43] Wenbo Zhang, Kaixuan Wang, Yutao Ouyang, Xiaoyu Huang, Liyang Li, Kailun Su, Weiyang Jin, Wenhao Chai, Haotian Liang, Zhiyang Dou, Yue Chen, and Tianxing Chen. An unexpected robot policy: Early evaluations of GPT-6 Astra on RoboDojo and beyond.arXiv preprint arXiv:2609.24170, 2026 on PithPith's reviewarXiv
- [44] Yixian Zhang, Huanming Zhang, Feng Gao, Xiao Li, Zhi- hao Liu, Yi Nie, Chunyang Zhu, Jiaxing Qiu, Yuchen Yan, Jiyuan Liu, Wenhao Tang, Jiaji Rao, Zhengru Fang, Changxu Wei, Yu Wang, Wenbo Ding, and Chao Yu. Harness VLA: Steering frozen VLAs into reliable manipulation primitive... on PithPith's reviewarXiv
- [45] Xufeng Zhao, Mengdi Li, Cornelius Weber, Muham- mad Burhan Hafez, and Stefan Wermter. Chat with the en- vironment: Interactive multimodal perception using large language models. InIEEE/RSJ Int. Conf. Intell. Robots Syst., pages 3590–3596, 2023 on Pith
- [46] Youwei Zhou, Bohan Hou, Yankai Fu, Yaoxu Lyu, Yifan Ye, Shilin Shan, et al. Large multimodal models for di- rect robot control: GPT-6 as a beginning.arXiv preprint arXiv:2610.28259, 2026. 10 ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models Supplem... on PithPith's reviewarXiv
- [48] Prefer small, deliberate motions near objects. Any motion whose wrist force exceeds {f_cap:g} N is aborted by the safety layer on Pith
- [49] When something is uncertain and cannot be seen, you may probe it physically (touch, lift, push) if your tools allow it on Pith
- [50] Be honest in done: report success=true only if you believe the task is actually complete; give a verdict when the task asks for one. ... Figure F1. The system prompt shared by all conditions, verbatim without its frames and units, its wrist-sensor line, and rules 1 and 5. Tabl... on Pith
- [9149] (a) Successes and (b) damaged handles out of 50. Daggers mark the rows of Table 5 of the main paper, lighter bars the variants, and hatched bars the oracle position script. On Lock a success also needs a correct decision. unseen. On CoM we see no leak, as it answers offset in ... 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