LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing

Boda Huo; Changyi Lin; Ding Zhao; Jie Tan; Mingyang Yu; Shiqi Liu; Tingnan Zhang; Wenhao Yu; Yikai Wang; Yiyue Luo

REVIEW 3 major objections 5 minor 70 references

LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LocoTouch: a quadruped learns, entirely in simulation, to carry loose unbalanced objects on its back using only touch, and the skill transfers to the real robot with no fine-tuning.

desk verdict A credible zero-shot tactile transport system on a quadruped; the core result stands, but the tactile sim fidelity and real-world breadth are thinner than the abstract implies. read the letter →

arxiv 2505.23175 v2 pith:MIO7Y2DG submitted 2025-05-29 cs.RO

Changyi Lin , Yuxin Ray Song , Boda Huo , Mingyang Yu , Yikai Wang , Shiqi Liu , Yuxiang Yang , Wenhao Yu

show 4 more authors

Tingnan Zhang Jie Tan Yiyue Luo Ding Zhao

This is my paper · ORCID

classification cs.RO

keywords tactilesensingquadrupedalrobotssim-to-realtransferloco-manipulationdistributedsensorreinforcementlearningteacher-studentdistillationobjecttransport

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a quadrupedal robot can perform sustained dynamic manipulation — carrying unsecured cylindrical objects on its back over long distances — using only tactile feedback, with a policy trained purely in simulation and deployed on real hardware without any fine-tuning. The authors build a piezoresistive sensor of 221 touch cells (taxels) covering the robot's entire back, simulate it with a calibrated 'expanded collision' model that reproduces real contact spread, and distill a teacher policy that sees privileged object state into a student policy that sees only binary touch signals. If true, this would show that dense touch can replace explicit object-state sensing in contact-rich legged manipulation, and that faithful tactile simulation does not require expensive soft-body modeling. The reported evidence includes near-perfect simulation success rates, perfect transport of six diverse real objects over 6 meters, a 60-meter run with a slippery bottle, and stability under pushes and on uneven terrain.

What carries the argument

The load-bearing mechanism is the expanded collision model for tactile simulation: each simulated taxel is given a fixed enlarged collision area of 18.3 × 17.5 mm (the real sensing area is 14.3 × 12.8 mm), calibrated once by sliding a metal ball across the real sensor and recording when a contact activates one taxel versus two. Because a contact near a taxel's edge then activates several neighboring taxels, the model reproduces the coupled contact spread of the real soft sensing material without querying the position of each contact point, keeping massively parallel training fast. Two further mechanisms carry the argument: the two-stage teacher–student distillation, in which a PPO teacher with privileged object state is distilled by behavior cloning and DAgger into a student with a Conv-GRU tactile encoder and simulated 20–40 ms sensor latency; and the adaptive gait reward, whose symmetricity function $f_{\mathrm{sym}}$ scores one diagonal leg pair's swing-air-time history against the other pair's, promoting symmetric, low-frequency trotting without fixed timing references.

What would settle it

Press a calibrated ball or a flat-bottomed object onto the real sensor at several force levels spanning the training range and slide it across a row of taxels while recording the activated binary map; then replay the same trajectory in simulation with the 18.3 × 17.5 mm expanded collision model. The fidelity claim is falsified if the number and identity of activated taxels diverge systematically with force or contact pose — for example, if the real sensor activates a third taxel at high force where the simulation never does, or if edge contacts light different neighbors than the model predicts; in that case the zero-shot success would have to be credited to domain randomization and the policy's tolerance of sensor mismatch rather than to the simulated tactile map.

Watch

Extended reading notes

Core claim

The central claim is that a simulation-trained tactile policy can balance and transport unsecured objects on a quadruped's back in the real world with zero fine-tuning. The authors attribute this to three ingredients working together: a high-density distributed tactile sensor of 221 taxels (touch cells) covering 250 × 180 mm of the robot's back; an efficient simulation of that sensor in which each taxel's collision volume is enlarged to a calibrated 18.3 × 17.5 mm area so that a contact near a taxel's edge triggers neighboring taxels, reproducing the force-conditioned contact spread of the real piezoresistive skin; and a teacher–student pipeline in which a policy trained with privileged object-state input is distilled, via behavior cloning and DAgger (an interactive imitation algorithm), into a student that reads only a binary tactile map encoded by a convolutional-GRU network. A central design choice is the adaptive gait reward: instead of rewarding a fixed foot-contact schedule, it scores diagonal-pair synchrony and lateral-pair alternation through a symmetricity function of measured air times, so gait frequency emerges from task performance rather than an external timing reference. The claim is supported by simulation success rates (teacher 98.8–99.8%, student 95.8–96.7%, and 0% when tactile or object-state input is removed) and by real-world tests on a Unitree Go1 carrying objects from a 0.03 kg glue stick to a 1.26 m poster tube, including slippery drink bottles, over 60 meters and under external pushes.

Load-bearing premise

The load-bearing premise is that a single fixed calibration — enlarging every simulated taxel's collision area to 18.3 × 17.5 mm, measured once by sliding a ball across the real sensor and checked on just two object placements (Section 4, Appendix B.2) — reproduces the real sensor's force-dependent contact spread well enough for all the objects, weights, and contact poses the robot will meet, so the touch signals the policy was trained on still match reality at deployment.

Editorial extensions

If this is right

A student policy that sees only touch and proprioception (no vision, no object-state estimate) keeps 95.8–96.7% transport success over 10 seconds in simulation, versus 98.8–99.8% for the state-privileged teacher, and scores 5/5 real-world trials on every object tested — touch alone carries enough information for sustained balancing.
The one-time ball calibration of enlarged taxel areas is sufficient for zero-shot sim-to-real of a contact-rich legged skill, indicating that expensive soft-body or hydroelastic tactile simulation is not needed for this class of tasks.
Rewarding gait symmetry and alternation without timing references yields adaptive stepping: the robot slows its step frequency when the velocity command drops and speeds it back up when the object slides backward, while keeping the two diagonal pairs' air times matched.
The trained policy generalizes beyond its training distribution to objects from 0.03 to 1.45 kg, 0.10 to 1.26 m long and 0.03 to 0.18 m in diameter, including non-cylindrical cups and wrenches, and to slopes, gravel, and rough terrain never seen in training.
The Limitations section states that transport fails for cylinders whose axis aligns with the robot's forward direction, because training randomizes yaw only within ±30° and the torso is narrow side-to-side, and that performance degrades on stairs because training uses flat ground only.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension: replay a library of real contact trajectories — different objects, force levels, and contact poses — through the simulator and measure how often the simulated binary map matches the real one; if the match rate is high across that library, the fixed-area model rather than policy robustness is what explains the zero-shot transfer.
The inverse coupling the authors observe between robot pitch and object forward displacement, produced by a policy that never sees object state, suggests the student's tactile embeddings implicitly track the carried load's pose and momentum; probing the Conv-GRU features against motion-capture object states would test that internal-model hypothesis.
Adding vision, the extension the paper itself proposes, would most directly attack its two named failure modes — longitudinal cylinder poses and stairs — since both are cases where the binary touch map alone is arguably ambiguous about the object's configuration and the terrain.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read LocoTouch. Worth your time: it does something previous quadruped work doesn't — carries unsecured cylinders on the back using only onboard tactile feedback, no container, no fastening, no real-world adaptation, and shows it working on six objects with 5/5 each. The core architecture is clean: teacher with privileged object state, DAgger distillation to a 221-taxel tactile student, Conv-GRU encoder. The ablations in simulation are decisive: student without touch gets 0%, teacher without object state 0%, locomotion-only about 0%. That gives me confidence the tactile signal is what the policy is using.

The genuinely new pieces: the task itself, the matrix-style distributed back sensor, the expanded-collision model for coupled taxels, and the adaptive gait reward. The gait reward is a nice bit of work — no fixed timing reference, symmetry evaluated on swing phase, and Figures 7 and 8 show real gait differences and drift without it. The sensor fabrication in Appendix B is concrete and reproducible in principle.

Soft spots. First, the tactile simulation fidelity claim is thin. The expanded taxel area is calibrated with one metal ball sliding across the sensor, and Figure 6 validates on two static poses only. The simulation uses a fixed enlarged-area box and a 0.05 N threshold; it doesn't encode the force-dependent spread the paper itself describes in Figure 3(b). For heavy, soft, or curved objects, the real activation pattern could be systematically wider. The 5/5 real-world successes on a 1.45 kg cylinder and a 1.26 m poster tube are evidence the simulation is good enough under test conditions, but they don't establish "high-fidelity" across the claimed range. I'd call that a real limitation, not a fatal one — binary signals and controller robustness can absorb a lot, and zero-shot transfer would almost certainly fail if the simulation were grossly wrong.

Second, the real-world quantitative evidence is narrow: six objects at 0.3 m/s for 6 m on flat ground. The long-distance (60 m), uneven-terrain, and perturbation claims rest on a single demonstration each, plus video stills. Those belong in the paper but should be labeled demonstrations, not repeated-trial results. Third, no code or sensor calibration data — with a custom sensor and simulation model, that's what makes the result independently checkable. Fourth, the appendix honestly reports the yaw-aligned failure mode and stair limitations; good, but it means the abstract's "wide range" should be parsed carefully.

Bottom line: for a reader working in legged loco-manipulation or tactile sim-to-real, this is a useful, credible contribution. It deserves peer review; the right outcome is probably accept with requests to publish code and calibration data and expand the real-world evaluation. I'd bring it to reading group.

Referee Report

3 major / 5 minor

Summary. LocoTouch presents a quadrupedal system for transporting unsecured cylindrical objects on the robot's back, using a custom 221-taxel tactile sensor, a simulation model with an enlarged collision area per taxel, a two-stage teacher-student training pipeline, and an adaptive gait reward. The teacher policy is trained with privileged object state in simulation and distilled via BC and DAgger into a student policy that observes binary tactile maps. The paper reports high simulation success rates, zero-shot real-world transport of six objects with varied size, weight, and surface properties, a single 60 m run, qualitative terrain generalization, and ablations showing that removing tactile or object-state input leads to near-zero success in simulation.

Significance. If the central claim holds, this is a notable advance: it would be one of the first demonstrations of a purely simulation-trained tactile policy transferring zero-shot to a quadruped for sustained dynamic manipulation, replacing privileged object-state sensing. The paper's strengths include a concrete external benchmark in the form of real-world deployment, internally consistent simulation ablations with strong negative controls (0% success without tactile input or object state), a scalable sensor fabrication approach, and a gait reward that does not require predefined contact schedules. The primary open risk is the fidelity of the simulated tactile map, which underpins the entire student policy and is calibrated with only a single indenter and validated on two static poses.

major comments (3)

[Section 4 and Appendix B.2, Fig. 6] The load-bearing claim of high-fidelity tactile simulation rests on a fixed enlarged collision area (18.3 x 17.5 mm) calibrated with a single sliding ball and validated on two static poses, whereas the sensor's own mechanism described in Fig. 3(b) is force-dependent contact spread; this dependence is not captured by the binary threshold f_sim=0.05 N or the 0.5% random flips. Because the student policy is trained and DAgger-refined exclusively on this simulated map, a systematic mismatch for the heavier, softer, or longer real objects in Appendix E.2 would translate directly into a miscalibrated tactile-to-action mapping, and the 5/5 successes per object at 0.3 m/s over 6 m cannot by themselves validate the 'high-fidelity' claim or the 'wide range' generalization claim. Please add a quantitative real-vs-sim comparison across the tested objects and contact locations, and either a sensitivity analysis over taxel size/threshold or explicit randomization over these parameters during training.
[Section 6.2, Table 1 and Appendix E.2] The quantitative support for the central sim-to-real claim is asymmetric: the success-rate table is computed in the training environment, and the real-world evidence is six objects times five trials at a single speed and distance plus one 60 m run, with no real-world no-tactile or locomotion-only baseline. To substantiate the statements that LocoTouch 'reliably transports a wide range' of objects and 'remains robust over long distances,' the paper should report the number of trials for the 60 m run, a real-world control without tactile input, and ideally a small set of commanded speeds and turning trajectories rather than only the single 0.3 m/s straight-line condition.
[Section 5.3 and Appendix C.5] The adaptive gait reward is a stated contribution and is claimed to be essential for robust transport, but the symmetricity function in Eq. (2) and Appendix C.5 introduces several hand-tuned parameters (alpha_task, alpha_tol, alpha_1, alpha_2, alpha_3, f_ub, f_lb) with no sensitivity analysis. The comparison to a constant gamma_sym in Figures 7 and 8 is suggestive, but it does not show that the reported benefit is robust to these parameter choices; please include an ablation or parameter study for at least the most sensitive parameters.

minor comments (5)

[Section 5.3] Just before Eq. (2), the text defines 'gamma_sim' but the symbol used everywhere else and in the equation is 'gamma_sym'; please unify the notation.
[Appendix E.2] The text refers to 'Table 17' for the object properties, but the properties are presented in Figure 17; the reference and numbering should be corrected.
[Appendix C.1, Table 5] The observation noise table lists 'Joint Position' twice, with the second row using units rad/s; this row should be labeled 'Joint Velocity'.
[Appendix B.3 and Section 5.4] The latency calibration reports approximately '0.02, s' and Section 5.4 quotes a range of 0.02-0.04 s, while Table 10 uses a simulation delay range of 0.025-0.05 s; please reconcile these values.
[Appendix E.3] The non-cylindrical object demonstrations are reported without trial counts or quantitative success rates; please state explicitly whether these are single demonstrations or repeated trials.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is validated by real-world zero-shot transfer, an external benchmark, and the tactile simulation model is calibrated against independent real-sensor measurements.

full rationale

Walking the derivation chain, the main claims are (i) a tactile simulation model that reproduces real contact spread, and (ii) a tactile student policy that transfers zero-shot to a physical quadruped. Neither claim reduces to its inputs by construction. The expanded collision model is calibrated once by sliding a metal ball across the real sensor (Appendix B.2), and the resulting simulated tactile maps are then compared against independently recorded real signals for two static cylinder poses (Section 6.3, Figure 6). This is a genuine external check: the calibration does not determine the validation images. The student policy is trained by BC and DAgger to imitate a teacher that observes privileged object state, while the student observes only tactile maps and proprioception; the distillation target is not the same as the student's input, so the student's success is not definitionally guaranteed. The real-world experiments, including 5/5 successes on six diverse objects (Appendix E.2), 60 m transport, uneven terrain, and perturbations, are an external benchmark outside the training distribution. The only noticeable self-citation is [53] for the two-stage distillation procedure, but this is a methodological reference, not a load-bearing theorem or uniqueness claim, and the surrounding method is standard BC plus DAgger with an independent citation [50]. No fitted parameter is renamed as a prediction, no ansatz is smuggled in via self-citation, and no known result is merely relabeled. The simplified binary and fixed-spread tactile model could be a fidelity and generalization limitation, but that is a correctness risk, not circularity, because the final claim is judged on the physical robot rather than on the training simulator alone.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central zero-shot transfer claim rests on the fidelity of the tactile simulation and the sufficiency of binary tactile observations. The expanded-collision taxel area is a calibrated free parameter rather than a derived quantity; the 0.05 N threshold, 0.5% flip rate, and latency range are hand-set. The training further assumes that IsaacLab's dynamics model is accurate, that trotting is the only needed gait, and that DAgger distillation preserves performance. No new physical entities are introduced; the sensor is a hardware artifact, not a theoretical construct.

free parameters (5)

Expanded taxel collision area = 18.3 x 17.5 mm
Calibrated from a ball-sliding test on the real sensor (Appendix B.2); one fixed enlargement factor for all simulated taxels, used for every contact in training.
Binary contact force threshold f_sim = 0.05 N
Hand-set threshold for converting simulated contact force to binary signals; real-world threshold is a separate calibrated voltage change.
Tactile signal flip rate = 0.5%
Noise injection chosen to narrow the sim-to-real gap for binary tactile maps (Section 4).
Tactile latency range = 0.02-0.04 s
Measured with a force gauge (Appendix B.3) and applied as a random delay during student-policy training.
Symmetricity-function tuning parameters = alpha_tol = 0.2, alpha = 1.2, plus upper/lower bounds
Hand-tuned within the adaptive gait reward to balance swing durations; directly shapes the learned gait and affects transport performance.

assumptions (5)

domain assumption IsaacLab simulation faithfully represents Go1 robot dynamics and contacts.
Policies are trained only in this simulator with no real-world fine-tuning; any systematic simulation error would invalidate the zero-shot claim. Invoked throughout Section 5 and C.1.
domain assumption Binary 221-taxel maps contain enough information to recover object state for transport.
Student policy replaces the teacher's privileged object state with binary tactile signals plus proprioception; the paper assumes this observation set is sufficient for the task (Section 5.4).
ad hoc to paper A single fixed enlarged-area taxel model generalizes to all object shapes and contact locations.
The expanded collision model is calibrated once using a ball and then applied to all simulated contacts without adapting to object geometry; this is a fitted approximation, not a physical deformation model (Section 4, Appendix B.2).
domain assumption Trotting is a suitable and sufficient gait for unsecured object transport.
The gait reward explicitly promotes trotting; no comparison with other gaits is provided, so the policy is constrained to this gait family (Section 5.3).
domain assumption Teacher-student distillation preserves task success in the real world.
Zero-shot transfer assumes the tactile student policy trained to imitate the privileged teacher in simulation will behave correctly on the physical robot; Table 1 supports the simulation part.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing." pith.science (2026). https://pith.science/paper/MIO7Y2DG

@misc{pith2026250523175,
  author       = {Pith},
  title        = {Pith review of: LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MIO7Y2DG}},
  note         = {Machine review of arXiv:2505.23175}
}

read the original abstract

Quadrupedal robots have demonstrated remarkable agility and robustness in traversing complex terrains. However, they struggle with dynamic object interactions, where contact must be precisely sensed and controlled. To bridge this gap, we present LocoTouch, a system that equips quadrupedal robots with tactile sensing to address a particularly challenging task in this category: long-distance transport of unsecured cylindrical objects, which typically requires custom mounting or fastening mechanisms to maintain stability. For efficient large-area tactile sensing, we design a high-density distributed tactile sensor that covers the entire back of the robot. To effectively leverage tactile feedback for robot control, we develop a simulation environment with high-fidelity tactile signals, and train tactile-aware transport policies using a two-stage learning pipeline. Furthermore, we design a novel reward function to promote robust, symmetric, and frequency-adaptive locomotion gaits. After training in simulation, LocoTouch transfers zero-shot to the real world, reliably transporting a wide range of unsecured cylindrical objects with diverse sizes, weights, and surface properties. Moreover, it remains robust over long distances, on uneven terrain, and under severe perturbations.

Figures

Figures reproduced from arXiv: 2505.23175 by the authors.

**Figure 2.** Training pipeline of LocoTouch. The teacher policy is trained via reinforcement learning using [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Distributed tactile sensing for LocoTouch. (a) Our custom, high-density tactile sensor covers the entire [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (18 more)

**Figure 4.** Figure 4: Symmetricity score fsym for varying previous air times tprev. When tprev < t ′ prev, fsym increases linearly up to an upper bound, encouraging longer air times and a more stable gait. When tprev > t′ prev, fsym penalizes excessively long swing durations, promoting mor…

**Figure 5.** Figure 5: Transport object under time-varying forward ve [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Comparison of tactile maps generated by dif [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Comparison of stepping air time for diagonal foot pairs over training. Our method promotes low-frequency symmetric gaits. As described in Section 6.1, the robot rapidly adapts its stepping frequency in response to command changes or object sliding while maintaining ga…

**Figure 8.** Figure 8: Baseline teacher policy exhibits lateral drifting due to asymmetric deformation of the foot pairs; while [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Materials, cutting files, and fabrication procedures for building the tactile sensors. The illustrated [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

**Figure 10.** Figure 10: Hardware set-up for expanded collision model calibration. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

**Figure 11.** Figure 11: The ball activates one or two taxels during calibration. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]

**Figure 12.** Figure 12: Left: Hardware set-up for sensor latency calibration. Right: Plot of readings from the tactile sensor and the force gauge. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

**Figure 13.** Figure 13: Simulation environment for teacher policy training in IsaacLab. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]

**Figure 14.** Figure 14: Illustrations of the symmetricity function. [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]

**Figure 15.** Figure 15: System setup in the real world. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]

**Figure 17.** Figure 17: Dimensions and weight of the objects used [PITH_FULL_IMAGE:figures/full_fig_p023_17.png]

**Figure 16.** Figure 16: Objects used in our experiments. They have [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]

**Figure 18.** Figure 18: LocoTouch balances and transports everyday objects such as cups and wrenches. [PITH_FULL_IMAGE:figures/full_fig_p023_18.png]

**Figure 19.** Figure 19: LocoTouch generalizes to diverse uneven terrains, including slopes, gravel, and rough surfaces. [PITH_FULL_IMAGE:figures/full_fig_p024_19.png]

**Figure 20.** Figure 20: The robot is capable of transporting a slippery drink bottle over long distances while consistently [PITH_FULL_IMAGE:figures/full_fig_p024_20.png]

**Figure 21.** Figure 21: The locomotion policy trained without object interaction totally fails to transport cylindrical objects. [PITH_FULL_IMAGE:figures/full_fig_p024_21.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 53 canonical work pages

[1]

Kumar, Z

A. Kumar, Z. Fu, D. Pathak, and J. Malik. Rma: Rapid motor adaptation for legged robots. arXiv preprint arXiv:2107.04034, 2021

arXiv 2021
[2]

Agarwal, A

A. Agarwal, A. Kumar, J. Malik, and D. Pathak. Legged locomotion in challenging terrains using egocentric vision. In Conference on robot learning, pages 403–415. PMLR, 2023

work page 2023
[3]

G. B. Margolis and P. Agrawal. Walk these ways: Tuning robot control for generalization with multiplicity of behavior. In Conference on Robot Learning, pages 22–31. PMLR, 2023

2023
[4]

Y . Yang, T. Zhang, E. Coumans, J. Tan, and B. Boots. Fast and efficient locomotion via learned gait transitions. In Conference on Robot Learning, pages 773–783. PMLR, 2022

work page 2022
[5]

G. B. Margolis, G. Yang, K. Paigwar, T. Chen, and P. Agrawal. Rapid locomotion via reinforce- ment learning. The International Journal of Robotics Research, 43(4):572–587, 2024

work page 2024
[6]

Cheng, A

X. Cheng, A. Kumar, and D. Pathak. Legs as manipulator: Pushing quadrupedal agility beyond locomotion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5106–5112. IEEE, 2023

work page 2023
[7]

Y . Ji, G. B. Margolis, and P. Agrawal. Dribblebot: Dynamic legged manipulation in the wild. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5155–5162. IEEE, 2023

work page 2023
[8]

Huang, Z

X. Huang, Z. Li, Y . Xiang, Y . Ni, Y . Chi, Y . Li, L. Yang, X. B. Peng, and K. Sreenath. Creating a dynamic quadrupedal robotic goalkeeper with reinforcement learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2715–2722. IEEE, 2023

work page 2023

Show all 70 references

[9]

Z. Fu, X. Cheng, and D. Pathak. Deep whole-body control: learning a unified policy for manipulation and locomotion. In Conference on Robot Learning, pages 138–149. PMLR, 2023

2023
[10]

Portela, G

T. Portela, G. B. Margolis, Y . Ji, and P. Agrawal. Learning force control for legged manipulation. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages 15366– 15372. IEEE, 2024

2024
[11]

C. Lin, X. Liu, Y . Yang, Y . Niu, W. Yu, T. Zhang, J. Tan, B. Boots, and D. Zhao. Locoman: Ad- vancing versatile quadrupedal dexterity with lightweight loco-manipulators. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6877–6884. IEEE, 2024

2024
[12]

X. He, C. Yuan, W. Zhou, R. Yang, D. Held, and X. Wang. Visual manipulation with legs. In Conference on Robot Learning, pages 4218–4234. PMLR, 2025. 9

2025
[13]

Valsecchi, N

G. Valsecchi, N. Rudin, L. Nachtigall, K. Mayer, F. Tischhauser, and M. Hutter. Barry: a high- payload and agile quadruped robot. IEEE Robotics and Automation Letters, 8(11):6939–6946, 2023

2023
[14]

Amanzadeh, T

L. Amanzadeh, T. Chunawala, R. T. Fawcett, A. Leonessa, and K. A. Hamed. Predictive control with indirect adaptive laws for payload transportation by quadrupedal robots. IEEE Robotics and Automation Letters, 2024

2024
[15]

M. Liu, Z. Chen, X. Cheng, Y . Ji, R.-Z. Qiu, R. Yang, and X. Wang. Visual whole-body control for legged loco-manipulation. arXiv preprint arXiv:2403.16967, 2024

2024 arXiv
[16]

C. Yang, G. N. Sue, Z. Li, L. Yang, H. Shen, Y . Chi, A. Rai, J. Zeng, and K. Sreenath. Collaborative navigation and manipulation of a cable-towed load by multiple quadrupedal robots. IEEE Robotics and Automation Letters, 7(4):10041–10048, 2022

2022
[17]

W. J. Jose and H. Zhang. Bilevel learning for dual-quadruped collaborative transportation under kinematic and anisotropic velocity constraints. arXiv preprint arXiv:2412.08644, 2024

2024 arXiv
[18]

J. Kim, R. T. Fawcett, V . R. Kamidi, A. D. Ames, and K. A. Hamed. Layered control for cooperative locomotion of two quadrupedal robots: Centralized and distributed approaches. IEEE Transactions on Robotics, 39(6):4728–4748, 2023

2023
[19]

Cheng, K

X. Cheng, K. Shi, A. Agarwal, and D. Pathak. Extreme parkour with legged robots. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11443–11450. IEEE, 2024

2024
[20]

Zhuang, Z

Z. Zhuang, Z. Fu, J. Wang, C. Atkeson, S. Schwertfeger, C. Finn, and H. Zhao. Robot parkour learning. arXiv preprint arXiv:2309.05665, 2023

2023 arXiv
[21]

Rudin, D

N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In Conference on Robot Learning, pages 91–100. PMLR, 2022

2022
[22]

Hoeller, N

D. Hoeller, N. Rudin, D. Sako, and M. Hutter. Anymal parkour: Learning agile navigation for quadrupedal robots. Science Robotics, 9(88):eadi7566, 2024

2024
[23]

K ¨aslin, H

R. K ¨aslin, H. Kolvenbach, L. Paez, K. Lika, and M. Hutter. Towards a passive adaptive planar foot with ground orientation and contact force sensing for legged robots. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2707–2714. IEEE, 2018

2018
[24]

X. A. Wu, T. M. Huh, A. Sabin, S. A. Suresh, and M. R. Cutkosky. Tactile sensing and terrain-based gait control for small legged robots. IEEE Transactions on Robotics, 36(1):15–27, 2019

2019
[25]

E. A. Stone, N. F. Lepora, and D. A. Barton. Walking on tactip toes: A tactile sensing foot for walking robots. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9869–9875. IEEE, 2020

2020
[26]

N. D. W. Mudalige, E. Nazarova, I. Babataev, P. Kopanev, A. Fedoseev, M. A. Cabrera, and D. Tsetserukou. Dogtouch: Cnn-based recognition of surface textures by quadruped robot with high density tactile sensors. In 2022 IEEE 95th Vehicular Technology Conference:(VTC2022- Spring...

2022
[27]

Vangen, T

A. Vangen, T. Barnwal, J. A. Olsen, and K. Alexis. Terrain recognition and contact force estimation through a sensorized paw for legged robots. In 2023 21st International Conference on Advanced Robotics (ICAR), pages 605–612. IEEE, 2023

2023
[28]

G. Shi, C. Yao, X. Liu, Y . Zhao, Z. Zhu, and Z. Jia. Foot vision: A vision-based multi-functional sensorized foot for quadruped robots. IEEE Robotics and Automation Letters, 2024. 10

2024
[29]

Z. Song, C. Li, Z. Quan, S. Mu, X. Li, Z. Zhao, W. Jin, C. Wu, W. Ding, and X.-P. Zhang. Tactid: High-performance visuo-tactile sensor-based terrain identification for legged robots. IEEE Sensors Journal, 2024

2024
[30]

L. Zhan, Y . Cao, Q. Chen, H. Guo, J. Gao, Y . Luo, S. Guo, G. Zhou, and J. Gong. Enable natural tactile interaction for robot dog based on large-format distributed flexible pressure sensors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 12493–...

2023
[31]

S. Guo, L. Zhan, Y . Cao, C. Zheng, G. Zhou, and J. Gong. Touch-and-heal: Data-driven affective computing in tactile interaction with robotic dog.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 7(2):1–33, 2023

2023
[32]

K. Yu, Y . Han, Q. Wang, V . Saxena, D. Xu, and Y . Zhao. Mimictouch: Leveraging multi-modal human tactile demonstrations for contact-rich manipulation. arXiv preprint arXiv:2310.16917, 2023

2023 arXiv
[33]

Guzey, B

I. Guzey, B. Evans, S. Chintala, and L. Pinto. Dexterity from touch: Self-supervised pre-training of tactile representations with robotic play. arXiv preprint arXiv:2303.12076, 2023

2023 arXiv
[34]

Pattabiraman, Y

V . Pattabiraman, Y . Cao, S. Haldar, L. Pinto, and R. Bhirangi. Learning precise, contact-rich manipulation through uncalibrated tactile skins. arXiv preprint arXiv:2410.17246, 2024

2024 arXiv
[35]

Huang, Y

B. Huang, Y . Wang, X. Yang, Y . Luo, and Y . Li. 3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing. arXiv preprint arXiv:2410.24091, 2024

2024 arXiv
[36]

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. arXiv preprint arXiv:2503.02881, 2025

2025 arXiv
[37]

Zhang, S

H. Zhang, S. Hu, Z. Yuan, and H. Xu. Doglove: Dexterous manipulation with a low-cost open-source haptic force feedback glove. arXiv preprint arXiv:2502.07730, 2025

2025 arXiv
[38]

Z.-H. Yin, B. Huang, Y . Qin, Q. Chen, and X. Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023

2023 arXiv
[39]

Y . Yuan, H. Che, Y . Qin, B. Huang, Z.-H. Yin, K.-W. Lee, Y . Wu, S.-C. Lim, and X. Wang. Robot synesthesia: In-hand manipulation with visuotactile sensing. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6558–6565. IEEE, 2024

2024
[40]

Z. Xue, H. Zhang, J. Cheng, Z. He, Y . Ju, C. Lin, G. Zhang, and H. Xu. Arraybot: Reinforcement learning for generalizable distributed manipulation through touch. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16744–16751. IEEE, 2024

2024
[41]

J. Yin, H. Qi, J. Malik, J. Pikul, M. Yim, and T. Hellebrekers. Learning in-hand translation using tactile skin with shear and normal force sensing. arXiv preprint arXiv:2407.07885, 2024

2024 arXiv
[42]

Y . S. Narang, K. Van Wyk, A. Mousavian, and D. Fox. Interpreting and predicting tactile signals via a physics-based and data-driven framework. arXiv preprint arXiv:2006.03777, 2020

2006 arXiv
[43]

Narang, B

Y . Narang, B. Sundaralingam, M. Macklin, A. Mousavian, and D. Fox. Sim-to-real for robotic tactile sensing via physics-based simulation and learned latent projections. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6444–6451. IEEE, 2021

2021
[44]

Kappassov, J.-A

Z. Kappassov, J.-A. Corrales-Ramon, and V . Perdereau. Simulation of tactile sensing arrays for physical interaction tasks. In 2020 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), pages 196–201. IEEE, 2020. 11

2020
[45]

Kasolowsky and B

U. Kasolowsky and B. B¨auml. Fine manipulation using a tactile skin: Learning in simulation and sim-to-real transfer. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13120–13127. IEEE, 2024

2024
[46]

D. P. Leins, F. Patzelt, and R. Haschke. Hydroelastictouch: Simulation of tactile sensors with hydroelastic contact surfaces. arXiv preprint arXiv:2501.08077, 2025

2025 arXiv
[47]

Ding, Y .-Y

Z. Ding, Y .-Y . Tsai, W. W. Lee, and B. Huang. Sim-to-real transfer for robotic manipulation with tactile sensory. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6778–6785. IEEE, 2021

2021
[48]

L. Yang, B. Huang, Q. Li, Y .-Y . Tsai, W. W. Lee, C. Song, and J. Pan. Tacgnn: Learning tactile-based in-hand manipulation with a blind robot using hierarchical graph neural network. IEEE Robotics and Automation Letters, 8(6):3605–3612, 2023

2023
[49]

Schulman, F

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

2017 arXiv
[50]

S. Ross, G. Gordon, and D. Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artifi- cial intelligence and statistics, pages 627–635. JMLR Workshop and Conference Proc...

2011
[51]

Sundaram, P

S. Sundaram, P. Kellnhofer, Y . Li, J.-Y . Zhu, A. Torralba, and W. Matusik. Learning the signatures of the human grasp using a scalable tactile glove. Nature, 569(7758):698–702, 2019

2019
[52]

Mittal, C

M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, et al. Orbit: A unified simulation framework for interactive robot learning environments. IEEE Robotics and Automation Letters, 8(6):3740–3747, 2023

2023
[53]

Y . Yang, G. Shi, C. Lin, X. Meng, R. Scalise, M. G. Castro, W. Yu, T. Zhang, D. Zhao, J. Tan, et al. Agile continuous jumping in discontinuous terrains. arXiv preprint arXiv:2409.10923, 2024

2024 arXiv
[54]

K. K. Patterson, I. Parafianowicz, C. J. Danells, V . Closson, M. C. Verrier, W. R. Staines, S. E. Black, and W. E. McIlroy. Gait asymmetry in community-ambulating stroke survivors.Archives of physical medicine and rehabilitation, 89(2):304–310, 2008

2008
[55]

W. Yu, G. Turk, and C. K. Liu. Learning symmetric and low-energy locomotion. ACM Transactions on Graphics (TOG), 37(4):1–12, 2018

2018
[56]

Abdolhosseini, H

F. Abdolhosseini, H. Y . Ling, Z. Xie, X. B. Peng, and M. Van de Panne. On learning symmetric locomotion. In Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games, pages 1–10, 2019

2019
[57]

Z. Su, X. Huang, D. Ordo˜nez-Apraez, Y . Li, Z. Li, Q. Liao, G. Turrisi, M. Pontil, C. Semini, Y . Wu, et al. Leveraging symmetry in rl-based legged locomotion control. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6899–6906. IEEE, 2024

2024
[58]

Mittal, N

M. Mittal, N. Rudin, V . Klemm, A. Allshire, and M. Hutter. Symmetry considerations for learning task symmetric robot policies. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7433–7439. IEEE, 2024

2024
[59]

Hutter, C

M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V . Tsounis, J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, et al. Anymal-a highly mobile and dynamic quadrupedal robot. In 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 38–44...

2016
[60]

Bledt, M

G. Bledt, M. J. Powell, B. Katz, J. Di Carlo, P. M. Wensing, and S. Kim. Mit cheetah 3: Design and control of a robust, dynamic quadruped robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2245–2252. IEEE, 2018. 13 A Related Work o...

2018
[61]

The definition of mirror symmetry depends heavily on robot morphology

Robot-specific assumptions. The definition of mirror symmetry depends heavily on robot morphology. For example, ETH Anymal [59] is symmetric both left-right and front-rear, while MIT Cheetah 3 robots [ 60] and Unitree Go1 are only left-right symmetric. This requires a careful ...
[62]

These methods often assume that robot dynamics and sensor configurations share the same symmetry as kinematics

Mismatch between kinematic and physical symmetry. These methods often assume that robot dynamics and sensor configurations share the same symmetry as kinematics. In practice, real robots rarely satisfy this assumption due to asymmetries in hardware, sensors, or payload
[63]

High-dimensional state challenges. Enforcing symmetry over the entire state–action space becomes increasingly difficult as observation spaces grow, particularly for policies that incorporate high-dimensional sensory inputs such as images
[64]

These methods promote symmetry only in kinematic trajectories, overlooking temporal aspects such as balanced swing and stance durations across legs

Neglect of temporal symmetry. These methods promote symmetry only in kinematic trajectories, overlooking temporal aspects such as balanced swing and stance durations across legs. Our key insight is that robots should exhibit symmetric gait behaviors when task commands (e.g., v...
[65]

Attach the conductive fabric onto the transfer tape
[66]

Fiber laser would be ideal since it doesn’t cut through the transfer tape

Laser cut the rows or columns on the conductive fabric, carefully tuning laser power and speed to avoid cutting through the transfer tape. Fiber laser would be ideal since it doesn’t cut through the transfer tape
[67]

Laser cut the outer shape of the sensor on both the conductive fabric and the transfer tape
[68]

Remove the excess conductive fabric, leaving only the patterned rows or columns adhered to the transfer tape
[69]

Attach external cables to one end of the rows and columns for readout
[70]

object dangerous state

Laser cut the Velostat layer and stack it with the two conductive layers prepared above. For additional resources on piezoresistive tactile sensors, we refer readers to the open-source tactile- skin toolkit. B.2 Expanded Collision Model Calibration We design a calibration tool...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Kumar, Z

A. Kumar, Z. Fu, D. Pathak, and J. Malik. Rma: Rapid motor adaptation for legged robots. arXiv preprint arXiv:2107.04034, 2021

arXiv 2021

[2] [2]

Agarwal, A

A. Agarwal, A. Kumar, J. Malik, and D. Pathak. Legged locomotion in challenging terrains using egocentric vision. In Conference on robot learning, pages 403–415. PMLR, 2023

work page 2023

[3] [3]

G. B. Margolis and P. Agrawal. Walk these ways: Tuning robot control for generalization with multiplicity of behavior. In Conference on Robot Learning, pages 22–31. PMLR, 2023

2023

[4] [4]

Y . Yang, T. Zhang, E. Coumans, J. Tan, and B. Boots. Fast and efficient locomotion via learned gait transitions. In Conference on Robot Learning, pages 773–783. PMLR, 2022

work page 2022

[5] [5]

G. B. Margolis, G. Yang, K. Paigwar, T. Chen, and P. Agrawal. Rapid locomotion via reinforce- ment learning. The International Journal of Robotics Research, 43(4):572–587, 2024

work page 2024

[6] [6]

Cheng, A

X. Cheng, A. Kumar, and D. Pathak. Legs as manipulator: Pushing quadrupedal agility beyond locomotion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5106–5112. IEEE, 2023

work page 2023

[7] [7]

Y . Ji, G. B. Margolis, and P. Agrawal. Dribblebot: Dynamic legged manipulation in the wild. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5155–5162. IEEE, 2023

work page 2023

[8] [8]

Huang, Z

X. Huang, Z. Li, Y . Xiang, Y . Ni, Y . Chi, Y . Li, L. Yang, X. B. Peng, and K. Sreenath. Creating a dynamic quadrupedal robotic goalkeeper with reinforcement learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2715–2722. IEEE, 2023

work page 2023