Pith. sign in

REVIEW 4 major objections 6 minor 48 references

TAC-LOCO: Unified Whole-Body Control for Quadrupedal TACtile-Informed LOCO-Manipulation

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Tactile-array feedback, encoded into a compact latent and fused with proprioception in one whole-body policy, lets a legged robot regulate grip online during loco-manipulation: object retention rises from 46.4% to 99.9% in simulation and ha

desk verdict The sim ablation cleanly isolates the tactile contribution and the core idea is new, but the abstract overstates the hardware evidence and the zero-shot claim leans on an unvalidated tactile sim-to-real map. read the letter →

arxiv 2607.10132 v2 pith:U7WBKM77 submitted 2026-07-11 cs.RO

Muqun Hu , Yuhao Zhou , Kabir Ray Malik , Chi Lin , Won Suk Lee , Yu She , Yan Gu This is my paper
classification cs.RO
keywords tactilesensingloco-manipulationwhole-bodycontrolreinforcementlearninggraspregulationsim-to-realtransferleggedrobotsobjectslip
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 argues that a working tactile sense at the gripper—not just proprioception—is what lets a legged robot carry an object through unknown, time-varying forces without either dropping it or crushing it. The authors train a single reinforcement-learning policy that controls legs, arm, and gripper together, feeding it a compact latent code of per-finger tactile pressure arrays alongside the usual joint and velocity readings. In 8,000 randomized simulated trajectories the tactile policy keeps the object for the entire run 99.9% of the time, while the proprioception-only version finishes only 46.4% of runs; it also cuts average grasping force by about 47%. The same policy, transferred without finetuning to a physical quadruped with a compliant tactile gripper, succeeds in 9 of 10 loco-manipulation trials against 5 of 10 for a fixed-grip baseline. If correct, this establishes tactile sensing as a load-bearing input for whole-body dynamic manipulation, not just for tabletop gripping.

What carries the argument

The load-bearing object is the tactile latent: each per-finger normal-force map is normalized frame-by-frame, flattened, and compressed through a two-layer neural network into a 32-dimensional vector that enters the policy's observation alongside proprioception and a history encoder. The policy itself is a single network with a shared backbone and separate leg and arm action heads, trained with on-policy reinforcement learning and a scheduled advantage-mixing mechanism that couples locomotion and manipulation objectives. A grasp-stability reward with three terms—force moderation, slip penalty, and retention—shapes the online grip regulation. Sim-to-real transfer rests on per-frame normalizat

What would settle it

In the same 8,000-rollout evaluation, set the tactile latent to a fixed constant while keeping everything else identical; the paper's claim predicts object retention should fall from 99.9% to roughly 46.4%. If the policy still completes most rollouts, tactile input is not the mechanism driving the reported advantage.

Watch

Extended reading notes

Core claim

The paper's central claim is that in-hand tactile array readings, compressed into a 32-dimensional latent and joined with proprioception, let a single whole-body policy continuously regulate grasp force while walking: it tightens the grip when the object begins to slip under rising external load and relaxes when the load is suddenly removed. The authors demonstrate this in simulation, with the tactile policy holding the object for the whole trajectory in 99.9% of 8,000 randomized rollouts versus 46.4% without tactile input, and with 47% lower average grasping force than a policy trained without the force-moderation reward. On hardware, the same policy succeeds zero-shot in 9 of 10 dynamic lo

Load-bearing premise

Simulated tactile maps, after per-frame normalization and domain randomization, carry the same relative contact-state information as the real compliant finger pads, so behaviors learned in simulation transfer zero-shot to hardware.

Editorial extensions

If this is right

  • A proprioception-only whole-body policy cannot reliably prevent slip under time-varying object loads; direct tactile observation is what closes the gap (46.4% vs 99.9% retention in simulation).
  • Grasp force can be moderated by learning rather than prescribed: removing the force-moderation term leaves success unchanged but roughly doubles average gripping force.
  • Dense slip and retention rewards are necessary for learning; without them the policy must rely on sparse episode termination and succeeds less often.
  • Adaptive grip regulation transfers zero-shot from simulation to hardware, so normalized synthetic tactile maps carry enough contact-state information for deployment.
  • Unseen object shapes can be handled if the policy responds to contact-distribution changes rather than memorized geometry.

Reading between the lines

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

  • One implication the authors leave implicit: because the policy reads only relative, normalized contact signals, tactile calibration—mapping taxels to absolute force units—may be unnecessary for adaptive grasping, which would simplify deploying the same approach on other compliant finger sensors.
  • The sudden-release curriculum models any detachable constraint, so the same framework should extend to harvesting, unplugging, and assembly tasks where a resisting force disappears abruptly; this is testable by substituting the spring-magnet release with a snap-fit connector.
  • A natural stress test: lower the tactile sampling rate or add temporal delay to the latent; if success degrades steeply, the policy is truly slip-reactive, not just passively stiff.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes TAC-LOCO, a unified reinforcement-learning framework for quadrupedal loco-manipulation in which high-dimensional tactile array readings from compliant gripper fingers are encoded into a compact latent and combined with proprioception to control legs, arm, and gripper jointly. The policy is trained in Isaac Lab with randomized spring-damper external forces and sudden-release events, using grasp-stability rewards that penalize excessive contact force, object slip, and object-EE displacement. The main claims are that (i) tactile feedback enables online grasp adaptation under time-varying loads, (ii) tactile observations materially improve slip-free completion in simulation (99.9% vs 46.4% over 8,000 randomized rollouts), and (iii) zero-shot hardware deployment on a Unitree Go2/Interbotix WidowX 250 achieves 90% success on loco-manipulation versus 50% for a fixed-grip Deep WBC baseline, with generalization to unseen objects and perturbations.

Significance. If the central claims hold, the paper makes a useful contribution: it demonstrates that tactile-array feedback can be a load-bearing input modality for whole-body loco-manipulation, rather than a quasi-static manipulation add-on. The large-sample simulation ablation (§5.2) is a genuine strength: it compares tactile versus no-tactile observations under an otherwise common reward and architecture, and the 46.4% vs 99.9% gap over 8,000 rollouts is compelling evidence that the observation channel matters inside the simulator. The paper also ships concrete implementation details (reward table, architecture dimensions, environment randomization ranges), which supports reproducibility. The hardware experiments, while small (n=10 per setting), include zero-shot transfer, unseen-object generalization, and manual perturbation tests, giving external evidence that the learned behavior is not purely a simulation artifact. However, the hardware baseline comparison is confounded, and the zero-shot transfer rests on an unquantified assumption about normalized simulated versus real tactile maps. These issues need to be addressed before the central cross-domain claim is fully supported.

major comments (4)
  1. [§6.5 / Table 4] The hardware baseline comparison does not isolate the tactile channel. Deep WBC is compared with (a) no tactile observations, (b) no learned gripper action (fixed width), and (c) a different whole-body policy training framework. The 0.9 vs 0.5 difference could therefore be caused by adaptive gripper actuation alone, by the whole-body policy difference, or by tactile feedback. The simulation ablation in §5.2 does isolate tactile observations, but only inside the simulator. To support the claim that tactile feedback is load-bearing on hardware, an additional hardware baseline is needed: the same TAC-LOCO policy with tactile observations removed (or masked), while keeping the same gripper action space and rewards. Without this, the abstract and contribution (c) overstate what the hardware experiment demonstrates.
  2. [§4.1 / App. A.2 / §7] The zero-shot transfer claim hinges on the assumption that normalized SDF/Kelvin-Voigt normal-force maps are informationally equivalent, up to scale/relative dynamics, to real FlexiTac readings during incipient slip and grip-width changes. The only evidence is the single static qualitative comparison in Fig. 12(b). No distributional metric is reported under varying grasp force, slip, or grip-width variation, and the authors themselves concede in §7 that the policy can infer relative changes in contact state but not enforce an exact absolute grasp-force threshold. This is more than a peripheral caveat: if real maps differ in contact spread, hysteresis, or force-response shape, the learned latent-to-grip mapping may not be what drives the reported hardware behavior. Please add a quantitative sim-to-real tactile-map comparison under representative contact-state changes (e.g., varying normal
  3. [Abstract / §1 contribution (b) / Tab. 2 and Fig. 4] The abstract's headline numbers — '47% reduction in grasping force and an object drop rate of less than 1%' — are simulation results (Tab. 2 for force reduction, Fig. 4 for drop rate). The abstract and contribution (b) present them without a simulation qualifier. On hardware, the reported success rates are 1.0 and 0.9 over 10 trials per task, and no hardware grasp-force measurement is reported. Please move the simulation-only numbers into the simulation section or explicitly mark them as simulation results in the abstract and contributions.
  4. [§5.2 / Fig. 4] The proprioception-only baseline is described as 'does not receive tactile observations but still observes the gripper width through joint position feedback.' This is the right control for isolating tactile input in simulation, and it is a strength. However, the figure reports only the fraction of rollouts retaining the object over trajectory progress. The authors should also report the mean and variance of trajectory progress at failure for both conditions, and state whether the 46.4% vs 99.9% comparison uses the same randomized seed sets. This would strengthen the claim that the gap is not an artifact of a single large force-release condition dominating the baseline failures.
minor comments (6)
  1. [§3] Typo: 'can be can be handled by upstream perception' — duplicate 'can be'.
  2. [§4.1 / Eq. (1)] The spring-damper external force model is a key part of the task definition, but the representation of p_e,t could be clarified: it is the vector from the object to the initial equilibrium position, and the unit vector appears in the damping term. Please define the sign convention explicitly.
  3. [App. A.1 / A.2] The normalization reference scale F_ref and contact threshold τ_F are not given numerical values in the main text or appendix. Please report the values used, since they affect the tactile observation and the domain-randomization story.
  4. [Tab. 2] The column header 'Width (mm)↑' is ambiguous — it appears to report average gripper width, and '↑' presumably indicates that higher width is better (less force). Please clarify in the caption or use a more descriptive header.
  5. [Fig. 12] The caption says 'Comparison of tactile patterns obtained in simulation and real-world experiments,' but the comparison is qualitative and shows only a static grasp. Consider a quantitative metric (e.g., normalized taxel activation maps or contact-area overlap) and at least one dynamic condition.
  6. [§6.2 / §6.3] The generalization and disturbance experiments are reported narratively without trial counts or success/failure breakdowns. Please report the number of trials per unseen object and the number of perturbation trials, even if all succeeded, so the reader can assess the strength of the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the central tactile-vs-proprioception ablation and zero-shot hardware comparison stand independently; the acknowledged sim-to-real tactile mismatch is a transfer-validity limitation, not a definitional reduction.

full rationale

I walked the paper's derivation chain and found no step that reduces by construction to its own inputs. The main simulation result (Sec. 5.2) compares a tactile-informed policy against a proprioception-only baseline trained under the same reward structure and randomized dynamics; the same object-retention, slip, and force metrics are used for both. Since the baseline is optimized with the same rewards, the comparison genuinely tests whether the tactile observation channel helps the policy maximize that objective, rather than predicting a metric from a fitted input. The force-reduction and retention numbers reported in the abstract and Sec. 1(b) are trained reward outcomes (r_force, r_slip, r_retention in Sec. 4.3 and Table 9), not out-of-sample predictions; a reward term being optimized and then reported as an achieved objective is standard RL reporting, not a circularity. The hardware deployment (Sec. 6) is a zero-shot test against an external baseline, Deep WBC [10], which is independent of the authors' prior work; although the hardware baseline is confounded by fixed gripper width and different training, confounds are experimental-design concerns, not circular reasoning. The paper's own limitation in Sec. 7 explicitly concedes that after normalization the policy can infer relative contact-state changes but cannot enforce an exact absolute grasp-force threshold; this is an honest statement of the sim-to-real assumption, and it does not make the simulation or deployment evidence circular. Self-citations in the related-work sections (e.g., [15], [26], [41]) are contextual and not load-bearing; no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in through a self-citation. The tactile sim-to-real equivalence concern raised by the skeptical reader is a validity threat about whether normalized SDF/Kelvin-Voigt force maps preserve real contact patterns, but the paper does not define its target result in terms of that assumption; it is an unvalidated transfer assumption, not a circular step.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central claim rests on hand-set reward weights, hand-set task distributions, a spring-damper external-force abstraction, and a sim-to-real tactile assumption the authors themselves qualify. Because the success metrics (retention, force, slip) are the same quantities the reward terms optimize (Tab. 9), the sim numbers largely measure whether PPO optimized its own objective; the independent grounding is the observation ablation (§5.2) and the hardware generalization results (§6). No new physical entities are postulated; the tactile latent is a network-internal representation with no external evidence path.

free parameters (6)
  • Force-moderation reward weight w(r_force) = -0.005 × Σ‖F_b‖ per step
    Hand-set weight penalizing total gripper-object contact force (Tab. 9). It directly produces the headline '47% grasp-force reduction' via the Tab. 2 ablation (13.11 N vs 24.76 N); the number is a tuneable consequence of this weight, not an externally measured quantity.
  • Grasp-stability reward weights (r_retention=5.0, r_slip=−0.3, r_force=−0.005) = 5.0; −0.3; −0.005
    Hand-chosen (Tab. 9). Sim success is measured against object retention — the same quantity r_retention optimizes, so the evaluation metrics coincide with the optimization objectives.
  • Tactile normalization reference scale F_ref and contact threshold τ_F = F_ref numerically unspecified; τ_F 'near-zero'
    Per-frame normalization (F/F_max, with F/F_ref fallback below τ_F, clipped [0,1]) discards absolute calibration; F_ref is never given a value (§4.1), so exact tactile magnitudes are not reproducible from the text.
  • External-force distribution (spring-damper + release threshold) = k∈[0,18] N/m, c∈[0.1,0.5] N/(m·s⁻¹), F_thr∈[0.5,10] N
    Hand-set randomization ranges (Tab. 5) for the spring-damper model Eq. (1); they define the load ramps and sudden-release events the policy is trained for and shape the difficulty of both sim and hardware results.
  • Tactile contact-model randomization ranges = compliant stiffness [5,100]; penetration offset [1,3] mm; k_n=1.0, k_d=3×10⁻³
    Domain-randomization ranges (App. A.2) around TacSL constants, chosen by hand to cover sim-to-real spread. The transfer claim rests on these bounds containing reality.
  • Episode termination delay after object loss = 2.0 s
    Hand-set (§4.1); avoids training on post-failure states and thereby shapes the learned slip-recovery behavior.
assumptions (6)
  • domain assumption Isaac Lab/PhysX Kelvin–Voigt compliant contact (k_n=1.0, k_d=3e-3) approximates real elastomer-object interaction closely enough after normalization
    §4.1, App. A.1. Load-bearing for sim-to-real transfer; the paper's own §7 admits the simulator 'does not perfectly reproduce the real-world deformation of the gripper and tactile array.'
  • domain assumption Per-frame normalized tactile maps (F/F_max, clipped to [0,1]) preserve the relative contact/slip information the deployment policy needs
    §4.1, App. A.2. The entire adaptive-grasp mechanism depends on this; §7 concedes absolute grasp-force thresholds cannot be enforced.
  • ad hoc to paper Spring-damper restoring force with random release threshold (Eq. 1) is representative of real dynamic loads (e.g., fruit-stem detachment)
    §3-4.1. The task framing and sim success claims are defined with respect to this model; hardware springs (6.1/12.3 N/m) and magnets (2-5 N) were chosen to match it.
  • domain assumption PPO with advantage mixing (Eq. 4 of [10]) converges to a whole-body policy with valid leg/arm credit assignment
    App. B Eq. (4). Training technique taken from [10]; the paper adds a linear β schedule but does not analyze convergence properties.
  • domain assumption The proprioception-only baseline is a fair ablation isolating the tactile observation
    §5.2. Baseline lacks tactile input but retains gripper-width feedback; whether the 53-point gap stems from tactile information content vs. extra network capacity or optimization difficulty is not controlled (no added-capacity baseline).
  • domain assumption Hardware trials (n=10, reduced command subset) are representative of the simulated capability
    App. C admits the real-world evaluation 'uses a subset of the simulated end-effector trajectories'; small-n success rates 1.0/0.9/0.5 are reported without confidence intervals.
invented entities (1)
  • 32-dim tactile latent z_t (two-layer MLP encoder)
    purpose: Compresses 768-dim normalized tactile arrays into the policy observation to convey grip/slip state at deployment
    A learned network-internal representation. No checkpoints are released and no falsifiable handle exists outside this paper's own rollouts; the paper's only evidence of its adequacy is the sim/hardware behavior shown here. Not a physical postulate, but a new internal entity whose informativeness is asserted rather than externally verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TAC-LOCO: Unified Whole-Body Control for Quadrupedal TACtile-Informed LOCO-Manipulation." pith.science (2026). https://pith.science/paper/U7WBKM77

@misc{pith2026260710132,
  author       = {Pith},
  title        = {Pith review of: TAC-LOCO: Unified Whole-Body Control for Quadrupedal TACtile-Informed LOCO-Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7WBKM77}},
  note         = {Machine review of arXiv:2607.10132}
}
read the original abstract

Dynamic loco-manipulation requires legged robots to coordinate whole-body motion while maintaining stable physical interaction with grasped objects under uncertain external forces. While tactile sensing has been widely studied for robotic manipulation, its role in dynamic whole-body control remains largely unexplored. Existing works without tactile feedback commonly grasp firmly rather than regulate the grasp according to the interaction. We propose TAC-LOCO, a tactile-augmented unified reinforcement learning framework that encodes tactile array observations from compliant grippers into a compact latent representation and joins it with proprioception for unified control of the legs, arm, and gripper. With effective grasp stability reward design, the policy learns to simultaneously track body velocity and end-effector trajectories, moderate grasp force, and prevent object slip under both gradual load changes and sudden release events. We deploy the policy zero-shot on a Unitree Go2 with an Interbotix WidowX 250 arm and tactile gripper, demonstrating dynamic tactile-informed loco-manipulation under varying external interactions, achieving a 47% reduction in grasping force and an object drop rate of less than 1%.

Figures

Figures reproduced from arXiv: 2607.10132 by the authors.

Figure 1
Figure 1. We present a unified tactile-informed whole-body control policy for dynamic loco￾manipulation, reducing grasping force and preventing object slip during unknown and time-varying external forces. The policy can accommodate a variety of object shapes and sizes. 1 Introduction Legged robots have rapidly evolved from robust locomotion platforms [1, 2] into mobile manipula￾tion systems capable of interacting with complex… view at source ↗
Figure 2
Figure 2. Framework of TAC-LOCO. We encode high-dimensional tactile array information into [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Grasping width changes during object-EE relative velocity and time-varying external force. We first evaluate TAC-LOCO in simula￾tion to study how tactile feedback affects grasp regulation during dynamic loco￾manipulation. We design experiments to answer the following questions: (1) Can the learned policy adapt the gripper width in response to changing external forces and EE motion (Sec. 5.1)? (2) Does tactile feedba… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Fraction of rollouts maintaining grasp contact with the object over trajectory progress. Tactile-informed policy substantially improves slip-free completion of the trajectory. Tactile feedback improves robustness under dynamic loco-manipulation. We next evaluate whethe…
Figure 5
Figure 5. Figure 5: Simulation evaluation under time-varying external forces with a wider workspace. Each [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Distributions of the linear-velocity, angular-velocity, and end-effector position tracking [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Ratio of environments terminated over policy update iterations. Slip and retention re￾wards improve task success. Policy Variant Avg. Force (N) ↓ Width (mm) ↑ TAC-LOCO full reward 13.11 24.65 w/o Force Reg. Reward 24.76 22.20 w/o Retention/Slip Reward 25.74 21.62 [PIT…
Figure 8
Figure 8. Figure 8: Real-world evaluation tasks. (a) Manipulation under time-varying external forces, where [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Grasp regulation behavior of the trained policy. (a) The artificial strawberry is pulled [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Real-world robustness to unmodeled object disturbances. During loco-manipulation, ran [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Benchmarking against [3]. The baseline uses a fixed gripper width and therefore cannot actively compensate for changes in the external force. The top row shows gradual in-hand object slip during loco-manipulation. The bottom row shows a failure case in which the objec…
Figure 12
Figure 12. Figure 12: Tactile simulation settings and sim-to-real comparison. (a) Effects of the elastomer stiff [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 12
Figure 12. Figure 12: Tactile simulation settings and sim-to-real comparison. (a) Effects of the elastomer stiff [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 6 linked inside Pith

  1. [41]

    P. Zhou, Y . Zhou, Q. Luu, S. Han, H. Zhang, B. Huang, Y . Li, A. Ajoudani, Z. Xu, and Y . She. Learning tactile-aware quadrupedal loco-manipulation policies.arXiv preprint arXiv:2604.27224, 2026

    arXiv 2026

  2. [1]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter. Learn- ing agile and dynamic motor skills for legged robots.Science robotics, 4(26):eaau5872, 2019

    2019

  3. [2]

    Iqbal, Y

    A. Iqbal, Y . Gao, and Y . Gu. Provably Stabilizing Controllers for Quadrupedal Robot Loco- motion on Dynamic Rigid Platforms.IEEE Trans. Mechatron., 25(4):2035–2044, Aug. 2020

  4. [3]

    Zhuang, Z

    Z. Zhuang, Z. Fu, J. Wang, C. G. Atkeson, S. Schwertfeger, C. Finn, and H. Zhao. Robot parkour learning. In7th Annual Conference on Robot Learning, 2023. URLhttps: //openreview.net/forum?id=uo937r5eTE

    2023

  5. [4]

    Iqbal, S

    A. Iqbal, S. Veer, and Y . Gu. Analytical solution to a time-varying lip model for quadrupedal walking on a vertically oscillating surface.Mechatron., 96:103073, 10 2023

    2023

  6. [5]

    Y . Gao, V . Paredes, Y . Gong, Z. He, A. Hereid, and Y . Gu. Time-varying foot placement control for humanoid walking on swaying rigid surface.IEEE Trans. Rob., 2025

    2025

  7. [6]

    Z. He, S. Teng, T.-Y . Lin, M. Ghaffari, and Y . Gu. Invariant filtering for full-state estimation of ground robots in noninertial environments.IEEE/ASME Trans. Mechatron., 2025, in press

    2025

  8. [7]

    Misenti, B

    S. Misenti, B. Hertel, B. Weng, R. Donald, A. Jawaji, M.-T. Kosoko-Thoroddsen, J. G. Trafton, A. Norton, R. Azadeh, and Y . Gu. Experimental evaluation of commercial quadruped robots: stability and performance in non-inertial environments.Int. J. Intel. Rob. Appl., 9(3):1109– 1134, 2025

    2025

Show all 48 references
  1. [8]

    L. C.-Y . Tay, I. Chang, and Y . Gu. Hybridmimic: Hybrid rl-centroidal control for humanoid motion mimicking.arXiv preprint arXiv:2603.06775, 2026

    2026

  2. [9]

    Mandali, Z

    F. Mandali, Z. He, and Y . Gu. Proprioceptive invariant state estimation for humanoid robots on noninertial ground.IEEE/ASME Trans. Mechatron., 2026, in press

    2026

  3. [10]

    Z. Fu, X. Cheng, and D. Pathak. Deep whole-body control: Learning a unified policy for manipulation and locomotion. In6th Annual Conference on Robot Learning, 2022. URL https://openreview.net/forum?id=zldI4UpuG7v

    2022

  4. [11]

    S. R. Iyer, I. Chang, A. Z. Liu, Y . Gu, Z. Kingston, et al. Vectorizing projection in manifold-constrained motion planning for real-time whole-body control.arXiv preprint arXiv:2604.13323, 2026

    2026 arXiv

  5. [12]

    M. Hu, W. Chen, W. Li, F. Mandali, Z. He, R. Zhang, P. Krisna, K. Christian, L. Benaharon, D. Ma, K. Ramani, and Y . Gu. Pace: Physics augmentation for coordinated end-to-end rein- forcement learning toward versatile humanoid table tennis.arXiv preprint arXiv:2509.21690, 2026

    2026

  6. [13]

    Y . Gu, G. Shi, F. Shi, I.-C. Chang, Y .-J. Wang, Q. Cheng, Z. Olkin, I. Lopez-Sanchez, Y . Feng, J. Zhang, H. Su, and K. Sreenath. Evolution of humanoid locomotion control. 2025, under review

    2025

  7. [14]

    Chang, X

    I.-C. Chang, X. Huang, T.-Y . Lin, S. Teng, W. Li, M. Ghaffari, J. Yi, Y . Gu, et al. A sur- vey of legged robotics in non-inertial environments: Past, present, and future.arXiv preprint arXiv:2604.20990, 2026

    2026 arXiv

  8. [15]

    C. Lin, Y . R. Song, B. Huo, M. Yu, Y . Wang, S. Liu, Y . Yang, W. Yu, T. Zhang, J. Tan, Y . Luo, and D. Zhao. Locotouch: Learning dynamic quadrupedal transport with tactile sensing. In9th Annual Conference on Robot Learning, 2025. URLhttps://openreview.net/forum?id= XaJkbK02Vm

    2025

  9. [16]

    M. Liu, Z. Chen, X. Cheng, Y . Ji, R.-Z. Qiu, R. Yang, and X. Wang. Visual whole-body control for legged loco-manipulation. In8th Annual Conference on Robot Learning, 2024. URLhttps://openreview.net/forum?id=cT2N3p1AcE. 13

    2024

  10. [17]

    X. Liu, B. Ma, C. Qi, Y . Ding, N. Xu, Zhaxizhuoma, G. Zhang, P. Chen, K. Liu, Z. Jia, C. Guan, Y . Mo, J. Liu, F. Gao, J. Zhong, B. Zhao, and X. Li. Mlm: Learning multi-task loco-manipulation whole-body control for quadruped robot with arm.IEEE Robotics and Automation Letters...

    2026

  11. [18]

    Zhang, H

    H. Zhang, H. Yu, L. Zhao, A. Choi, Q. Bai, Y . Yang, and W. Xu. Learning multi-stage pick- and-place with a legged mobile manipulator.IEEE Robotics and Automation Letters, 10(11): 11419–11426, 2025. doi:10.1109/LRA.2025.3608425

    2025

  12. [19]

    M. Jung, J. Lee, Z. He, D. Youm, J. Mun, H. Kim, H. Oh, D. Choi, J. Hur, J. Song, and J. Hwangbo. Learning dynamic pick-and-place for a legged manipulator.IEEE Robotics and Automation Letters, 11(6):7652–7659, 2026. doi:10.1109/LRA.2026.3688092

    2026

  13. [20]

    Portela, G

    T. Portela, G. B. Margolis, Y . Ji, and P. Agrawal. Learning force control for legged manip- ulation. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 15366–15372. IEEE, 2024

    2024

  14. [21]

    R.-Z. Qiu, Y . Song, X. Peng, S. A. Suryadevara, G. Yang, M. Liu, M. Ji, C. Jia, R. Yang, X. Zou, et al. Wildlma: Long horizon loco-manipulation in the wild. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 10011–10019. IEEE, 2025

    2025

  15. [22]

    D. Hou, C. Zhu, Z. Zhang, Z. Li, C. Guo, and Y . Liu. Efficient learning of a unified policy for whole-body manipulation and locomotion skills. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5455–5461. IEEE, 2025

    2025

  16. [23]

    G. Pan, Q. Ben, Z. Yuan, G. Jiang, Y . Ji, S. Li, J. Pang, H. Liu, and H. Xu. Roboduet: Learning a cooperative policy for whole-body legged loco-manipulation.IEEE Robotics and Automation Letters, 10(5):4564–4571, 2025. doi:10.1109/LRA.2025.3551230

    2025

  17. [24]

    P. Zhi, P. Li, J. Yin, B. Jia, and S. Huang. Learning a unified policy for position and force control in legged loco-manipulation. In9th Annual Conference on Robot Learning, 2025. URLhttps://openreview.net/forum?id=MpJTyAqA0t

    2025

  18. [25]

    B. Xu, H. Weng, Q. Lu, Y . Gao, and H. Xu. FACET: Force-adaptive control via impedance reference tracking for legged robots. In9th Annual Conference on Robot Learning, 2025. URL https://openreview.net/forum?id=EcOGafgvuC

    2025

  19. [26]

    Y . Zhou, W. S. Lee, Y . Gu, and Y . She. Tactile-reactive gripper with an active palm for dexter- ous manipulation.npj Robotics, 4:13, 2026

    2026

  20. [27]

    F. R. Hogan, J. Ballester, S. Dong, and A. Rodriguez. Tactile dexterity: Manipulation primi- tives with tactile feedback. In2020 IEEE International Conference on Robotics and Automa- tion (ICRA), pages 8863–8869. IEEE, 2020

    2020

  21. [28]

    Y . Zhou, P. Zhou, S. Wang, and Y . She. In-hand singulation and scooping manipulation with a 5 dof tactile gripper. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5238–5243. IEEE, 2024

    2024

  22. [29]

    Sleiman, F

    J.-P. Sleiman, F. Farshidian, M. V . Minniti, and M. Hutter. A unified mpc framework for whole-body dynamic locomotion and manipulation.IEEE Robotics and Automation Letters, 6 (3):4688–4695, 2021. doi:10.1109/LRA.2021.3068908

    2021

  23. [30]

    Chiu, J.-P

    J.-R. Chiu, J.-P. Sleiman, M. Mittal, F. Farshidian, and M. Hutter. A collision-free mpc for whole-body dynamic locomotion and manipulation. In2022 International Conference on Robotics and Automation (ICRA), pages 4686–4693, 2022. doi:10.1109/ICRA46639.2022. 9812280. 14

    2022

  24. [31]

    A. Rigo, M. Hu, S. K. Gupta, and Q. Nguyen. Hierarchical optimization-based control for whole-body loco-manipulation of heavy objects. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 15322–15328, 2024. doi:10.1109/ICRA57147.2024. 10611656

    2024

  25. [32]

    Molnar, J

    L. Molnar, J. Cheng, G. Fadini, D. Kang, F. Zargarbashi, and S. Coros. Whole-body inverse dynamics mpc for legged loco-manipulation.IEEE Robotics and Automation Letters, 11(1): 898–905, 2026. doi:10.1109/LRA.2025.3636005

    2026

  26. [33]

    Cheng, D

    J. Cheng, D. Kang, G. Fadini, G. Shi, and S. Coros. Rambo: Rl-augmented model-based whole-body control for loco-manipulation.IEEE Robotics and Automation Letters, 10(9): 9462–9469, 2025. doi:10.1109/LRA.2025.3594984

    2025

  27. [34]

    Y . Ma, A. Cramariuc, F. Farshidian, and M. Hutter. Learning coordinated badminton skills for legged manipulators.Science robotics, 10(102):eadu3922, 2025

    2025

  28. [35]

    Zhang, Y

    Y . Zhang, Y . Yuan, P. Gurunath, I. Gupta, S. Omidshafiei, A.-a. Agha-mohammadi, M. Vazquez-Chanlatte, L. Pedersen, T. He, and G. Shi. Falcon: Learning force-adaptive hu- manoid loco-manipulation.arXiv preprint arXiv:2505.06776, 2025

    2025

  29. [36]

    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, 9(7): 6720–6727, 2024

    2024

  30. [37]

    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, 24(16):26487–26495, 2024

    2024

  31. [38]

    E. A. Stone, N. F. Lepora, and D. A. Barton. Walking on tactip toes: A tactile sensing foot for walking robots. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9869–9875. IEEE, 2020

    2020

  32. [39]

    Murooka, T

    M. Murooka, T. Hoshi, K. Fukumitsu, S. Masuda, M. Hamze, T. Sasaki, M. Morisawa, and E. Yoshida. Tact: Humanoid whole-body contact manipulation through deep imitation learning with tactile modality.IEEE Robotics and Automation Letters, 10(8):7819–7826, 2025. doi: 10.1109/LRA.2...

    2025

  33. [40]

    Y . Niu, Z. Fang, B. Chen, S. Zhou, R. Senthilkumaran, H. Zhang, B. Chen, C. Qiu, H. E. Tseng, J. Francis, et al. Learning versatile humanoid manipulation with touch dreaming.arXiv preprint arXiv:2604.13015, 2026

    2026 arXiv

  34. [42]

    W. Yuan, S. Dong, and E. H. Adelson. Gelsight: High-resolution robot tactile sensors for estimating geometry and force.Sensors, 17(12):2762, 2017

    2017

  35. [43]

    Akinola, J

    I. Akinola, J. Xu, J. Carius, D. Fox, and Y . Narang. Tacsl: A library for visuotactile sensor simulation and learning.IEEE Transactions on Robotics, 41:2645–2661, 2025. doi:10.1109/ TRO.2025.3547267

    2025

  36. [44]

    Huang, Y

    B. Huang, Y . Wang, X. Yang, Y . Luo, and Y . Li. 3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing. In8th Annual Conference on Robot Learning, 2024. URLhttps: //openreview.net/forum?id=bk28WlkqZn

    2024

  37. [45]

    Huang, J

    B. Huang, J. Xu, I. Akinola, W. Yang, B. Sundaralingam, R. O’Flaherty, D. Fox, X. Wang, A. Mousavian, Y .-W. Chao, and Y . Li. VT-refine: Learning bimanual assembly with visuo- tactile feedback via simulation fine-tuning. In9th Annual Conference on Robot Learning,

  38. [46]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms, 2017. URLhttps://arxiv.org/abs/1707.06347

    2017 arXiv

  39. [47]

    Mittal, P

    M. Mittal, P. Roth, J. Tigue, A. Richard, O. Zhang, P. Du, A. Serrano-Mu ˜noz, X. Yao, R. Zurbr ¨ugg, N. Rudin, L. Wawrzyniak, M. Rakhsha, A. Denzler, E. Heiden, A. Borovicka, O. Ahmed, I. Akinola, A. Anwar, M. T. Carlson, J. Y . Feng, A. Garg, R. Gasoto, L. Gulich, Y . Guo, M...

    2025 arXiv

  40. [2025]

    URLhttps://openreview.net/forum?id=bOVF8Rj33i. 15

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.