Pith. sign in

REVIEW 5 major objections 6 minor 38 references

Making products among robot sensors explicit beats simply making the policy network bigger.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 21:24 UTC pith:SRDYFJAQ

load-bearing objection Solid robotics packaging of factorized poly proprio conditioning with real capacity controls; the big LIBERO jump is the soft pillar, not the Humanoid-Gym one. the 5 major comments →

arxiv 2607.23473 v1 pith:SRDYFJAQ submitted 2026-07-26 cs.RO

PRISM: Polynomial Representations for Interaction-Structured Motor Control

Seung Hyun Lee , Stella X. Yu This is my paper
classification cs.RO
keywords polynomial interactionmotor controlsensorless complianceproprioceptive conditioninghumanoid locomotioncontact-rich manipulationdiffusion policyembodied AI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Robot policies usually feed joint angles, velocities, and commands into a plain multilayer network and hope it invents the right couplings. Many quantities that actually matter for control—power, slip, contact impulse, compliance—are products of those signals, not the signals alone. This paper argues that those interactions should be built into the policy as a compact, learnable polynomial layer rather than left for the network to discover. The method, PRISM, forms a few latent factors and multiplies them element-wise so higher-order structure is available without listing every monomial. On humanoid walking and contact-rich arm tasks it outperforms both standard and larger-capacity networks under the same sensors and controllers, and it produces yielding contact behavior without force or tactile hardware. The practical claim is that interaction structure is not interchangeable with parameter count, so polynomial proprioceptive conditioning should be a default architectural choice in embodied motor control.

Core claim

Exposing low-degree factorized polynomial interactions among deployment-available physical observations improves motor policies beyond what matched-capacity multilayer networks achieve, and can induce sensorless compliance-like contact regulation without force, wrench, tactile input, contact labels, or an admittance controller. On humanoid locomotion and contact-rich manipulation, that architectural change—not extra width alone—raises tracking, survival, and task success while making mechanics-inspired quantities more linearly readable from the policy representation.

What carries the argument

PRISM: a factorized polynomial proprioceptive module. Two (or recursively K) learned affine maps of the deployable state form latent factors; element-wise products with a learned, near-zero-initialized gate yield ψ_K, which keeps a first-order path and adds compact higher-order interaction features before the unchanged downstream actor or diffusion conditioner.

Load-bearing premise

The paper assumes that the cues that matter for these tasks live mainly in low-degree products of signals the robot already has, so a compact polynomial layer is enough without force sensing, long history, or hidden contact geometry.

What would settle it

Train the same locomotion and LIBERO setups with a width-matched MLP versus PRISM under identical data, rewards, and controllers: if the larger MLP matches or beats PRISM on episode length, tracking error, survival, and success, and contact-force traces show no sensorless yielding advantage, the claim that interaction structure cannot be replaced by capacity collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Default robot actors and visuomotor conditioners should expose factorized polynomial proprioceptive interactions rather than only linear or MLP state embeddings.
  • Matched-capacity ablations become necessary when claiming an architectural inductive bias in motor control; width alone is not a substitute for an interaction basis.
  • Contact-rich policies can seek compliance-like force regulation from proprioceptive products without adding force/tactile sensors or admittance loops at deployment.
  • Linear probes on frozen policy features become a diagnostic for whether mechanics-inspired quantities (slip, power, impulse, work) are more accessible after the representation change.
  • The same module can drop into both PPO-style locomotion actors and diffusion or vision-language-action proprioceptive pathways without changing actions or low-level controllers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If low-degree products of proprioception are this useful, many existing sim-to-real stacks may be leaving easy structure on the table simply by keeping linear state encoders.
  • A natural next stress test is tasks where failure is dominated by unobserved geometry or materials: PRISM should plateau while sensor-augmented baselines keep improving.
  • Adaptive or input-dependent polynomial degree could matter when some joints need quadratic coupling and others stay nearly linear.
  • The sensorless compliance result suggests re-examining ‘force-free’ industrial insertion and wiping stacks that currently bolt on admittance only because the policy never saw multiplicative state structure.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes PRISM, a factorized elementwise polynomial module that exposes degree-K interactions among deployment-available proprioceptive variables while preserving the policy objective, action interface, and low-level controller. In PPO-based Humanoid-Gym locomotion, PRISM reportedly outperforms both a standard MLP and a parameter-matched larger MLP; in LIBERO imitation learning, replacing Diffusion Policy's linear proprioceptive conditioner with a degree-2 PRISM layer raises reported average success from 63.8% to 91.0%. Additional experiments test BFM-Zero and SmolVLA, ablate polynomial degree, probe linear recoverability of physical quantities, evaluate evaluation-time perturbations, and include a small SO-101 pilot. The authors also interpret lower contact-force behavior and factor ablations as evidence of sensorless, compliance-like interaction regulation.

Significance. If the results hold, the paper provides a useful and easily adopted architectural inductive bias: a compact, learnable interaction basis that requires no force, tactile, contact-label, or privileged deployment input. Particularly valuable are the parameter-count controls, degree ablation, evaluation on two architecturally stronger backbones, simulator-only post-hoc physical probes, and the preliminary real-robot pilot. The Humanoid-Gym comparison is promising evidence that interaction structure is not interchangeable with width. Establishing the LIBERO result at converged training budgets and the compliance result without success-conditioned selection would make the proposed standard architectural choice substantially more credible.

major comments (5)
  1. [§4.2 / Appendix A.2 / Tables 2, 9, 13–14] Tables 2, 9, and 13: the headline Diffusion Policy result (63.8% to 91.0%) is obtained after only 20K updates and 10 rollouts per task. No training-length comparison or convergence evidence is provided, and the companion experiment on the stronger SmolVLA backbone finds only +3.05 points at an aligned 80K checkpoint (Table 14), using seed 1000. Consequently, the present evidence cannot distinguish a genuine interaction-structure gain at convergence from a large short-training/regularization effect on an undertrained baseline. Because Table 2 anchors the manipulation and sensorless-compliance claims, the paper should report longer or converged task-specific Diffusion Policy training, ideally with learning curves and multiple seeds, and reconcile the result with the much smaller SmolVLA gain.
  2. [§4.2 / Appendix A.1 / Tables 1 and 8] Table 1 and Table 8 give inconsistent parameter counts for what appear to be the same Humanoid-Gym actors: Table 1 reports 0.926M for MLP and 1.321M for both Larger MLP and PRISM, while Table 8 reports 527K, 922K, and 922K, respectively, explicitly labeled as actor parameters. Since the capacity-control conclusion—that interaction structure rather than parameter count drives the gain—depends on exact matching, the authors must explain what additional modules are included in Table 1 or correct one table.
  3. [§4.2 / Appendix C.1 / Tables 1 and 18] The default degree and locomotion numbers are not internally consistent. §4.2 states that degree 2 is the default and degree 3 performs best, while Table 1's PRISM row reports episode length 2233.4 and survival 92.50. Under the stated same Humanoid-Gym setup, Table 18 reports PRISM-D2 as 2198.7/91.00 and PRISM-D3 as 2296.5/92.50. Table 1 therefore matches neither row exactly. Please identify which degree, checkpoint-selection rule, and seed protocol generated Table 1, and make the main result and degree ablation directly comparable.
  4. [§4.2 / Figure 4 / Appendix A.4] Appendix A.4 states that the contact diagnostics are success-conditioned, averaging only successful episodes. Since PRISM has a much higher success rate than Diffusion Policy and MCC-Sensorless, this selection rule can mechanically bias contact-force comparisons toward the method that succeeds more often; Figure 4 further relies on one representative rollout. The sensorless-compliance claim should instead be supported by contact-aligned force, impulse, power, and approach-speed distributions over all rollouts, or by comparisons stratified by success/failure and task, with confidence intervals.
  5. [§4 / Tables 1–3 and 13–15] Several key comparative tables do not quantify uncertainty or replication. Table 1 is described as averaged over five seeds but reports no standard errors; Tables 2 and 13 use only 10 rollouts per task; Table 14 reports one seed for SmolVLA; and Table 15 uses 20 rollouts per perturbation condition. For the paper's central cross-domain claim, report seed/task-level confidence intervals or significance tests and state the number of independently trained policies for every entry.
minor comments (6)
  1. [§1 / §3.2–3.3 / Fig. 2] §1 and Fig. 2: the introduction says PRISM is applied 'after' the MLP backbone in RL, whereas §3.2–3.3 and Fig. 2 describe polynomial conditioning of the proprioceptive input before the downstream actor. Please use one consistent description and block diagram.
  2. [§4.2 / Table 2] Table 2 calls MCC-Oracle a non-deployable force-access ablation that provides an upper bound, but its measured success (64.5%) is below PRISM and only slightly above Diffusion Policy. 'Force-access ablation' is accurate; 'upper bound' is not established by the result.
  3. [§4.3 / Table 4] Table 4 supports improved linear recoverability, but the title's phrase 'Newtonian quantities' is stronger than the operational targets shown; slip, contact impulse, and contact work are simulator-computed proxies. Consider 'mechanics-inspired quantities,' as used in the caption.
  4. [§4.3 / Table 5] Table 5's post-hoc factor names are useful, but the mapping from two affine branches' largest weights to a named dominant input product is qualitative. Please give the exact ranking/threshold rule and, if possible, show the leading weight products for at least one factor.
  5. [Appendix B.5 / Table 17] The SO-101 pilot is appropriately framed as preliminary, but Table 17 should include trial-level confidence intervals or otherwise avoid visual emphasis given only five demonstrations and ten trials per method.
  6. [§4.2 and throughout] Several typographical and formatting issues should be corrected, including 'Our approaches quickly before contact,' the missing space in the project URL sentence, and the inconsistent multiplication notation in Fig. 1 and Eq. (2).

Circularity Check

0 steps flagged

No significant circularity: empirical architecture paper with external training objectives and non-load-bearing post-hoc analysis.

full rationale

PRISM is proposed as a factorized polynomial proprioceptive conditioner (Eqs. 1–6) and evaluated by standard external objectives (PPO, DDPM noise prediction) against capacity-matched baselines. Reported gains (Tables 1–3) are empirical rollout metrics, not quantities defined by or fitted to equal the architecture. Linear probes (Table 4) and factor ablations (Table 5) use simulator-derived targets and post-hoc weight interpretations only for analysis; they are not training labels and do not force the success/survival claims. Citations to multiplicative/polynomial nets and Diffusion Policy are background inductive bias and backbone choice, not uniqueness theorems or self-citation chains that make the results true by construction. The derivation chain does not collapse any claimed prediction into its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The claim rests on a physics-motivated inductive bias (action-relevant cues live in low-order products of observables), standard RL/IL training assumptions, and several design knobs (degree, latent width, α initialization, gradual activation) chosen for stability and performance. No new physical particle or force is postulated; the invented object is the PRISM conditioner. Free parameters are architectural/training choices rather than constants fitted to prove a law.

free parameters (4)
  • polynomial degree K (default 2) = K=2 default; K=3 best in loco ablation
    Chosen by ablation; degree 3 is slightly better on Humanoid-Gym but degree 2 is defaulted for simplicity (Table 18, §4.2).
  • latent interaction width = 256
    Sets factor dimension for u_t, v_t and capacity match to larger MLP (256 in main Humanoid-Gym/Diffusion setups).
  • α_k initialization / scale = near 0; α2=1e-2 in BFM-Zero/SmolVLA
    α near zero (or 1e-2 with RMSNorm in stronger backbones) so the module starts nearly linear and interactions ramp up; hand-set for training stability (§3.2, Table 10).
  • MCC-Sensorless force-proxy gains (kf, delay, EMA, noise) = kf=0.015, delay=2, EMA=0.9, noise std=0.15
    Selected on a validation split for the compliance baseline comparison (Table 12); affects relative ranking vs PRISM.
axioms (5)
  • domain assumption Many action-relevant physical quantities (power, slip, contact impulse/work, inertial couplings) are polynomial/multiplicative in deployment-available signals even when not listed as inputs.
    Motivates the whole architecture (Intro, Fig. 1); plausible physics but not proven exhaustive for all failure modes.
  • ad hoc to paper A factorized elementwise construction ψ_k = ψ_{k-1} ⊙ (1 + α_k ⊙ (W_k x + b_k)) is an adequate compact substitute for explicit monomial bases.
    Core parameterization §3.2, inspired by cited multiplicative/polynomial nets but chosen as the paper's operational form.
  • domain assumption Width-scaled MLPs with matched parameter count are a sufficient control for 'capacity alone' versus interaction structure.
    Used to interpret Tables 1 and 3; alternative capacity axes (depth, attention, longer history) are not fully controlled.
  • domain assumption Standard PPO / DDPM / backbone training objectives remain valid when only the proprioceptive conditioner changes.
    §3.3–3.4; shared rewards, actions, and controllers across methods.
  • domain assumption Universal approximation does not imply efficient discovery of multiplicative structure from first-order inputs alone under practical data and width.
    Stated against Cybenko/Hornik citations in Intro; underpins why explicit products help.
invented entities (2)
  • PRISM factorized polynomial proprioceptive conditioner no independent evidence
    purpose: Expose learnable low-degree interaction features among robot observables inside existing RL and IL policy pipelines without new sensors.
    Primary proposed module (Eqs. 1–6); evaluated empirically rather than derived from a uniqueness theorem.
  • Emergent latent interaction factors (e.g., velocity-memory couplings named post hoc) no independent evidence
    purpose: Interpret which multiplicative directions drive actions after training (Table 5).
    Named by dominant input weights after ablation; not independently measured physical sensors.

pith-pipeline@v1.2.0-grok45-kimik3 · 18807 in / 3910 out tokens · 81531 ms · 2026-07-30T21:24:18.853037+00:00 · methodology

0 comments
read the original abstract

Robot policies are typically MLPs mapping observations to actions. Yet robot observations are physical variables, and many action-relevant cues arise not from individual variables but from their interactions; power, inertial effects, contact, slip, and compliance depend on products among observable signals. We introduce PRISM, a policy representation that makes polynomial interactions among observable physical variables explicit, learnable, and compact. Rather than listing all polynomial terms, PRISM uses a factorized polynomial module to expose higher-order interaction features efficiently. In reinforcement learning, it keeps the standard MLP backbone but applies a gradually activated element-wise polynomial function after it. In imitation learning, it replaces linear proprioceptive conditioning in Diffusion Policy with a polynomial layer trained end-to-end. Across humanoid locomotion and contact-rich manipulation, PRISM improves performance over standard MLP policies and larger MLPs with matched capacity, showing that interaction structure cannot be replaced by capacity alone. It also yields sensorless compliant behavior without force, wrench, tactile input, contact labels, or admittance control. These results suggest that polynomial representations should become a standard architectural choice for embodied motor control. The project page is available at https://lsh3163.github.io/prism/

Figures

Figures reproduced from arXiv: 2607.23473 by Seung Hyun Lee, Stella X. Yu.

Figure 1
Figure 1. Figure 1: Policies should expose multiplicative interactions among proprioception, actions, commands, and motion variables. We show representative rollout signals from locomotion tasks and manipulation tasks; each curve plots the corresponding physical quantity over rollout time. These crucial physical quantities are not directly provided as policy inputs, but they can be formed through higher-order interactions of … view at source ↗
Figure 2
Figure 2. Figure 2: PRISM exposes latent polynomial interactions within existing robot policy pipelines. In reinforcement learning, PRISM augments the actor with a polynomial interaction branch over deployment-available locomotion observations, while keeping the same action interface and low￾level PD controller. In imitation learning, PRISM replaces the standard linear proprioceptive con￾ditioning pathway with a degree-2 late… view at source ↗
Figure 3
Figure 3. Figure 3: PRISM improves both training stability and locomotion behavior. Locomotion re￾sults on Humanoid-Gym [10]. Left figure shows training curves with shaded standard error across seeds. Polynomial actors learn more reliably and reach higher performance than baselines. Right figure shows forward command-tracking traces under a fixed forward command, where the black line denotes the command and others denote real… view at source ↗
Figure 4
Figure 4. Figure 4: PRISM exhibits compliance-like contact regulation without force input. The top row visualizes a representative LIBERO [11] rollout for the turn on the stove task, with the con￾tact area highlighted in yellow. The bottom row plots contact force (Fc, orange) and end-effector speed (vEEF, blue), temporally aligned to the initial contact time t. While standard Diffusion Pol￾icy [9] also executes a fast approac… view at source ↗
Figure 5
Figure 5. Figure 5: Learned actor representations respond to unobserved dynamics shifts. t-SNE of the post-conditioner BFM-Zero actor features for reference motion 30. Each panel contains the same 30 matched samples per condition: nominal (green circles), low friction (blue triangles), and payload mass (orange squares). Features are standardized, reduced to at most 50 PCA dimensions, and projected with independently fitted t-… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 12 linked inside Pith

  1. [1]

    G. Cybenko. Approximation by superpositions of a sigmoidal function.Mathematics of con- trol, signals and systems, 2(4):303–314, 1989

    1989

  2. [2]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989

    1989

  3. [3]

    Cortes and V

    C. Cortes and V . Vapnik. Support-vector networks.Machine learning, 20(3):273–297, 1995

    1995

  4. [4]

    Sch ¨olkopf and A

    B. Sch ¨olkopf and A. J. Smola.Learning with kernels: support vector machines, regularization, optimization, and beyond. MIT press, 2002

    2002

  5. [5]

    S. M. Jayakumar, W. M. Czarnecki, J. Menick, J. Schwarz, J. Rae, S. Osindero, Y . W. Teh, T. Harley, and R. Pascanu. Multiplicative interactions and where to find them. InInternational conference on learning representations, 2020

    2020

  6. [6]

    A. Som, H. Choi, K. N. Ramamurthy, M. P. Buman, and P. Turaga. Pi-net: A deep learning ap- proach to extract topological persistence images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 834–835, 2020

    2020

  7. [7]

    Usevich, R

    K. Usevich, R. Borsoi, C. D ´erand, and M. Clausel. Identifiability of deep polynomial neural networks.Advances in Neural Information Processing Systems, 38:81809–81858, 2026

    2026

  8. [8]

    Maharaj and P

    S. Maharaj and P. B. Nair. Deep learning with learnable product-structured activations. InThe F ourteenth International Conference on Learning Representations

  9. [9]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 2024

    2024

  10. [10]

    Gu, Y .-J

    X. Gu, Y .-J. Wang, and J. Chen. Humanoid-gym: Reinforcement learning for humanoid robot with zero-shot sim2real transfer.arXiv preprint arXiv:2404.05695, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. Libero: Benchmarking knowl- edge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023

    2023

  12. [12]

    H. Shi, S. Hu, Y . Hou, W. Wang, K. Liu, and S. Song. Minimalist compliance control.arXiv preprint arXiv:2603.00913, 2026

    arXiv 2026

  13. [13]

    T. Xiao, W. Zhang, Y . Cheng, and J. Suo. Hope: High-order polynomial expansion of black- box neural networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46 (12):7924–7939, 2024

    2024

  14. [14]

    H. Li, Y . Chai, B. Lv, L. Ruan, H. Zhao, Y . Zhao, and J. Luo. Physics-informed neural network predictive control for quadruped locomotion.arXiv preprint arXiv:2503.06995, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    H. Hu, L. Feng, S. Chen, T. Zheng, D. Jiang, W. Chen, C. Zhang, G. Yang, and Y . Jin. Qui- etwalk: Physics-informed reinforcement learning for ground reaction force-aware humanoid locomotion under diverse footwear.arXiv preprint arXiv:2604.23702, 2026

    Pith/arXiv arXiv 2026

  16. [16]

    C. Guo, G. L’Erario, G. R. M. Leonori, M. Lorenzini, A. Ajoudani, and D. Pucci. Physics- informed learning for human whole-body kinematics prediction via sparse imus. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4472–

  17. [17]

    X. B. Peng, E. Coumans, T. Zhang, T.-W. Lee, J. Tan, and S. Levine. Learning agile robotic locomotion skills by imitating animals.arXiv preprint arXiv:2004.00784, 2020. 10

    Pith/arXiv arXiv 2004

  18. [18]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomo- tion over challenging terrain.Science robotics, 5(47):eabc5986, 2020

    2020

  19. [19]

    Zhang, N

    C. Zhang, N. Rudin, D. Hoeller, and M. Hutter. Learning agile locomotion on risky terrains. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11864–11871. IEEE, 2024

    2024

  20. [20]

    P. Zhi, P. Li, J. Yin, B. Jia, and S. Huang. Learning a unified policy for position and force control in legged loco-manipulation.arXiv preprint arXiv:2505.20829, 2025

    arXiv 2025

  21. [21]

    Kumar, Z

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

    Pith/arXiv arXiv 2021

  22. [22]

    Azulay, Z

    O. Azulay, Z. Xu, A. Scheffer, and S. X. Yu. Vigor: Visual goal-in-context inference for unified humanoid fall safety.arXiv preprint arXiv:2602.16511, 2026

    arXiv 2026

  23. [23]

    Lin and S

    K.-Y . Lin and S. X. Yu. Let humanoids hike! integrative skill development on complex trails. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22498– 22507, 2025

    2025

  24. [24]

    Zhang, Z

    X. Zhang, Z. Kou, C. Qin, M. Huang, E. Ristani, A. Kumar, L. Chen, K. He, A. Boularias, and L. Guan. Glove2hand: Synthesizing natural hand-object interaction from multi-modal sensing gloves.arXiv preprint arXiv:2603.20850, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Z. Liu, C. Chi, E. Cousineau, N. Kuppuswamy, B. Burchfiel, and S. Song. Maniwav: Learn- ing robot manipulation from in-the-wild audio-visual data.arXiv preprint arXiv:2406.19464, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    R. Wang, H. Geng, T. Li, F. Wang, G. Anumanchipalli, T. Darrell, B. Li, P. Abbeel, J. Malik, and A. A. Efros. The sound of simulation: Learning multimodal sim-to-real robot policies with generative audio.arXiv preprint arXiv:2507.02864, 2025

    arXiv 2025

  27. [27]

    X. Zhu, B. Huang, and Y . Li. Touch in the wild: Learning fine-grained manipulation with a portable visuo-tactile gripper.Advances in Neural Information Processing Systems, 38: 153783–153812, 2026

    2026

  28. [28]

    Jiang, H

    S. Jiang, H. Li, R. Ren, Y . Zhou, Z. Wang, and B. He. Kaiwu: A multimodal manipulation dataset and framework for robot learning and human-robot interaction.IEEE Robotics and Automation Letters, 2025

    2025

  29. [29]

    B. Xu, H. Weng, Q. Lu, Y . Gao, and H. Xu. Facet: Force-adaptive control via impedance reference tracking for legged robots.arXiv preprint arXiv:2505.06883, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    G. B. Margolis, M. Wang, N. Fey, and P. Agrawal. Softmimic: Learning compliant whole-body control from examples.arXiv preprint arXiv:2510.17792, 2025

    arXiv 2025

  31. [31]

    Schulman, F

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

    Pith/arXiv arXiv 2017

  32. [32]

    Shukor, D

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, et al. Smolvla: A vision-language-action model for afford- able and efficient robotics.arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Y . Li, Z. Luo, T. Zhang, C. Dai, A. Kanervisto, A. Tirinzoni, H. Weng, K. Kitani, M. Guzek, A. Touati, et al. Bfm-zero: A promptable behavioral foundation model for humanoid control using unsupervised reinforcement learning.arXiv preprint arXiv:2511.04131, 2025

    arXiv 2025

  34. [34]

    Cadene, S

    R. Cadene, S. Aliberts, F. Capuano, M. Aractingi, A. Zouitine, P. Kooijmans, J. Choghari, M. Russi, C. Pascal, S. Palma, et al. Lerobot: An open-source library for end-to-end robot learning.arXiv preprint arXiv:2602.22818, 2026. 11

    arXiv 2026

  35. [35]

    Kobayashi, C.-C

    M. Kobayashi, C.-C. Liao, K. Inoue, S. Yojima, and M. Takahashi. Motion capture dataset for practical use of ai-based motion editing and stylization.arXiv preprint arXiv:2306.08861, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Zhang and R

    B. Zhang and R. Sennrich. Root mean square layer normalization.Advances in neural infor- mation processing systems, 32, 2019

    2019

  37. [37]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014. 12 Appendix A Experimental Details Across all experiments, compared methods share observations, actions, controllers, data, objectives, and evaluation protocols. PRISM changes only the representation of deployment-available state inputs; it introduces ...

    Pith/arXiv arXiv 2014

  38. [38]

    Table 9: Shared Diffusion Policy configuration on LIBERO

    PRISM changes only the proprioceptive conditioner and is trained jointly with the policy, without auxiliary physical supervision. Table 9: Shared Diffusion Policy configuration on LIBERO. Item Setting Image size / observation steps128×128/2 Prediction horizon / action steps16/8 DDPM train / inference steps100/10 U-Net widths[128,256,512] Batch / optimizer...

    2048