Pith. sign in

REVIEW 1 major objections 6 minor 50 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

Robots gain on/off switch for physical contact

2026-07-10 01:56 UTC pith:EITB2BB5

load-bearing objection ContactMimic: contact-conditioned humanoid tracking that works, but per-motion training makes 'decoupling' easier than the headline implies the 1 major comments →

arxiv 2607.08742 v1 pith:EITB2BB5 submitted 2026-07-09 cs.RO

ContactMimic: Humanoid Object Interaction via Contact Control

Xinyao Li , Xialin He , Runpei Dong , Saurabh Gupta This is my paper
classification cs.RO
keywords contactcontactmimickeypointinteractionobjecttasksacrossaugmentation
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.

The paper identifies a specific failure mode in humanoid motion tracking: a robot can arrive at the correct body pose without actually touching the object it is supposed to interact with, rendering the motion useless for tasks like wiping a board or sitting on a chair. The authors argue that keypoint trajectories alone are an incomplete task specification because they encode geometry but not intent. Their solution, CONTACTMIMIC, augments a standard motion-tracking policy with a binary per-body-part contact command that is toggled at deployment time. The central technical problem is that human motion data entangles keypoints with contacts (a sitting motion almost always co-occurs with seat contact), so a policy trained naively would ignore the contact command and simply infer contact from the trajectory shape. The paper's contribution is a data augmentation pipeline that manufactures motion pairs sharing similar keypoint trajectories but opposing contact labels, forcing the policy to attend to the explicit contact command rather than shortcut through geometric cues. Combined with contact-aware rewards that penalize or encourage proximity to object surfaces, the resulting policy can, under the same keypoint trajectory, produce or suppress physical contact on command, enabling object manipulation without task-specific reward engineering.

Core claim

The paper demonstrates that decoupling contact behavior from keypoint geometry in a humanoid tracking policy requires manufactured training data where the same motion appears with both contact and no-contact labels. Without this augmentation, the policy ignores the contact command and infers contact from keypoints alone; with it, the policy learns to treat contact as an independent controllable variable, achieving measurable on/off contact control across diverse motions in both simulation and on real hardware.

What carries the argument

The load-bearing mechanism is a three-part trajectory augmentation scheme (contact-label flipping, object removal, and inflated-collision retargeting) that generates paired motions with similar keypoint trajectories but different contact labels. This is paired with two contact-aware reward terms: a label-matching reward (balanced accuracy or TP-FP) that rewards agreement between commanded and actual contact states, and a contact-distance reward that guides body parts toward or away from object surfaces based on the commanded label.

Load-bearing premise

The augmentation scheme is sufficient to break the keypoint-contact correlation in training data. If the manufactured motion pairs do not adequately cover the space of contact-independent motions, the policy could still shortcut by inferring contact from keypoints, which the no-augmentation ablation partially confirms but does not fully rule out for unseen motions or objects.

What would settle it

Deploy the trained policy on a novel motion where the keypoint trajectory strongly implies contact (e.g., a reaching motion toward a wall) but the contact command is set to suppress. If the policy makes contact anyway, the decoupling has failed and the policy is still inferring contact from geometry rather than obeying the command.

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

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

1 major / 6 minor

Summary. ContactMimic proposes augmenting humanoid keypoint-tracking policies with explicit, per-body-part binary contact labels that can be toggled at deployment time. The method has two components: (1) contact-aware rewards (contact label matching and contact distance, §3.1) that train the policy to follow commanded contacts, and (2) a trajectory augmentation scheme (contact-label flipping, object removal, inflated geometry, §3.2) that generates motion pairs sharing keypoint structure but differing in contact labels, breaking the shortcut where the policy infers contact from keypoints alone. The system is evaluated on 10 human-object interaction motions in simulation (Table 8, Fig. 4) and 5 motions on a real Unitree G1 (Table 2), with ablations on the augmentation scheme (Fig. 5, Table 9) and a comparison to the BeyondMimic keypoint-only baseline (Table 3). The central claim is that the policy 'successfully decouples contact behavior from keypoint geometry,' enabling contact-controllable loco-manipulation without task-specific rewards.

Significance. The problem is well-motivated: keypoint-only trackers can reproduce motion geometry while missing the contacts that make a task meaningful (wiping vs. waving). Exposing contact as a runtime-controllable input is a clean and general interface idea. The augmentation scheme to break keypoint-contact correlations is a concrete, falsifiable contribution, and the ablation in Fig. 5/Table 9 confirms its necessity. The sim-to-real transfer on 5 motions and the linear-probing analysis (Table 4) showing that proprioception encodes runtime contact state are valuable empirical contributions. The comparison to BeyondMimic (Table 3) demonstrates that contact conditioning yields higher contact metrics at comparable MPJPE, and the box-lifting result (Fig. 3) shows manipulation without task-specific rewards. These are meaningful steps toward purposeful humanoid-environment interaction.

major comments (1)
  1. The headline claim that the policy 'successfully decouples contact behavior from keypoint geometry' (Abstract, §1) is overstated relative to the evidence. All results come from per-motion policies (§5, Limitations), which is a much easier setting than decoupling across diverse motions. Even within this per-motion setting, Table 8 shows weak or absent controllability for several motions: Step foot on chair (T✔_near 0.61 vs T✘_near 0.60 contact bodies, ~2% reduction), Wipe whiteboard (0.65 vs 0.52, ~20%), and Lean on backrest I (1.38 vs 0.68, ~51%). The paper acknowledges that residual contact persists because 'the cost exceeds the contact-aware penalty under our reward weighting' (§F), but this means the policy has learned a reward-weighted compromise, not a true decoupling. The whiteboard-wiping motion—featured in Fig. 1, the abstract, and the real-world evaluation—is among the weakest.
minor comments (6)
  1. Table 3: The comparison to BeyondMimic is limited to a single baseline. Adding at least one more keypoint-tracking approach (e.g., ExBody2, OmniH2O) would strengthen the claim that keypoint-only control is insufficient.
  2. §3.1, Eq. (2): The balanced accuracy and TP-FP reward forms are both used, but the per-motion choice (Table 7) appears ad hoc. A principled criterion for when to use which would help.
  3. §3.2: The 1cm contact threshold for label extraction and the δ_infl ∈ [5,10]cm inflation range are stated but not justified. A sensitivity analysis on these parameters would strengthen the method.
  4. Figure 4: The arrow notation is compact but hard to parse. Consider adding a clearer legend or a companion table.
  5. Table 8 is large and sparse. Condensing or splitting it would improve readability.
  6. The project page URL (https://lixinyao11.github.io/contactmimic-page/) should be verified to be live at publication time, as the paper references video results.

Circularity Check

0 steps flagged

No circularity found: the policy is trained with contact rewards and evaluated on physics-measured contact metrics, with no self-definitional or fitted-input-as-prediction loops.

full rationale

The paper's central claim—that the policy decouples contact behavior from keypoint geometry and achieves contact controllability—is evaluated against externally measured quantities (contact bodies, contact impulse, object displacement) computed from simulation physics, not from the policy's own outputs. The contact labels used as policy input are extracted from retargeted motion data via geometric distance thresholds (§3.2), and the contact-aware rewards (§3.1) compare the actual contact state c_{t,b,p} against reference labels c̄_{t,b,p}. The evaluation trajectories T✔/T✘ (near/far) are constructed by flipping contact labels or inflating geometry, then the policy's resulting physical contact is measured independently. No step in the derivation chain reduces to its own inputs by construction. The linear probing result (Table 4) is presented as an analysis of learned representations, not as a load-bearing prediction. The per-motion training limitation and weak decoupling on some motions are correctness concerns, not circularity. The paper is self-contained against external benchmarks (BeyondMimic comparison, real-world deployment).

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, forces, or mathematical objects. The contact label map and augmentation strategies are methodological constructs, not postulated entities. The free parameters are standard RL reward shaping weights and thresholds, hand-tuned per the appendix. The axioms are domain assumptions standard in humanoid RL but worth surfacing as they bound the generality of the approach.

free parameters (8)
  • w_lm (contact label matching weight) = 4.0
    Hand-set reward weight for contact label matching term (Table 6).
  • w_cd (contact distance weight) = 3.0
    Hand-set reward weight for contact distance term (Table 6).
  • σ (contact distance Gaussian width) = 0.2m
    Shaping parameter for contact distance reward (Table 6).
  • δ (unwanted contact penalty threshold) = 0.05m
    Distance threshold for penalizing unwanted contacts (Table 6).
  • δ_infl (inflation offset) = [5,10]cm
    Range for object collision geometry inflation during augmentation (§D).
  • Contact threshold (label extraction) = 1cm
    Distance threshold for marking contact in retargeted trajectory (§3.2, §D).
  • λ (FPR penalty in TP-FP mode) = 1.0
    Penalty coefficient for false positive rate in sparse-contact reward (§3.1, §C).
  • p_rem, p_flip (augmentation probabilities) = 0.3, 0.2
    Per-episode probabilities for object removal and label flipping (Table 7).
axioms (4)
  • domain assumption Proprioception provides sufficient information to infer runtime contact state without contact sensors.
    The policy does not receive contact sensor input; the paper validates this with linear probing (Table 4) but it remains an assumption that holds for the tested motions and robot.
  • domain assumption Binary per-body-part contact labels are an adequate representation for specifying contact-rich tasks.
    The framework reduces contact to binary labels per body part (§3), ignoring contact force magnitude, location within a body part, and contact type (sliding, static). This is a modeling choice that limits expressiveness.
  • domain assumption HUMOTO dataset motions are representative enough of humanoid-relevant contact interactions.
    Training is limited to 10 clips from HUMOTO (§4.1); the diversity of interactions is bounded by this dataset.
  • domain assumption PhysX rigid-body simulation adequately models the contact dynamics for sim-to-real transfer.
    The policy is trained in Isaac Lab with PhysX (§4.1); sim-to-real transfer assumes the simulation contact model is sufficiently faithful.

pith-pipeline@v1.1.0-glm · 21668 in / 2944 out tokens · 343069 ms · 2026-07-10T01:56:00.842900+00:00 · methodology

0 comments
read the original abstract

Keypoint tracking alone is insufficient for object interaction tasks such as sitting on a chair, wiping a board, or pushing furniture, where the robot can reach the correct pose without making meaningful physical contact with the object. We present CONTACTMIMIC, a learning framework that tracks explicit partlevel binary contact commands alongside keypoint trajectories. CONTACTMIMIC is made possible through the use of contact-following rewards and a trajectory augmentation scheme aimed at breaking the correlations between keypoint trajectories and contact labels. The resulting policy successfully decouples contact behavior from keypoint geometry, and achieves precise physical contact as well as contact-controllability (produce or suppress contact during deployment as desired). Simulation experiments across 10 diverse human-object interaction motions confirm that CONTACTMIMIC exhibits contact controllability that enables it to complete manipulation tasks without task-specific rewards, while also outperforming keypoint-only trackers on contact-relevant tasks. Ablations confirm the necessity of the proposed trajectory augmentation scheme and sim2real deployment validates contact controllability in the real world across 5 different motions. Video results are available on https://lixinyao11.github.io/contactmimic-page/.

Figures

Figures reproduced from arXiv: 2607.08742 by Runpei Dong, Saurabh Gupta, Xialin He, Xinyao Li.

Figure 1
Figure 1. Figure 1: CONTACTMIMIC enables explicit contact control on a real humanoid across diverse interaction tasks. For each task, a same policy is commanded to either make the task-relevant contact (contact ✔) or to suppress it (contact ✘), by toggling a per-part contact label. Abstract: Keypoint tracking alone is insufficient for object interaction tasks such as sitting on a chair, wiping a board, or pushing furniture, w… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our pipeline. We retarget human-object interaction clips from HUMOTO, extracting both reference keypoint trajectories and per-body contact labels. We then synthesize aug￾mented motion pairs by inflated geometry, contact-label flipping and object removal. A contact￾conditioned policy is then trained on this data with contact-and-keypoint-tracking and contact-aware rewards, so that contact can be… view at source ↗
Figure 4
Figure 4. Figure 4: Per-motion visualization of contact controllability. Red arrows (near keypoints, contact ✔→✘, i.e. T ✔ near→T ✘ near) and blue arrows (far keypoints, contact ✔→✘, i.e. T ✔ far→T ✘ far) show contact metrics dropping with the contact command turned off. Both arrows confirm that the policy cor￾responds to the contact command. In the rightmost panel, the arrow beside each key joint marks whether its torque sho… view at source ↗
Figure 3
Figure 3. Figure 3: The box-lifting motion in simu￾lation. Top row: contact ✘, the same key￾points are tracked but the box is left un￾touched; bottom row: contact ✔, the robot grasps and lifts the box. Results. We plot the number of contacts and contact im￾pulses in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Removing proposed data augmentation hurts contact controllability. Solid arrows for our full method, are more consistently long and facing left (i.e. contacts reduce when commanded not to make contact) than the dotted arrows for the version trained without the augmentations from § 3.2. Motion Cha. Ref label Obs Layer 2 Wipe whiteboard 61 0.843 0.956 0.964 Sit in front of table 99 0.762 0.997 0.997 Lean on … view at source ↗
Figure 6
Figure 6. Figure 6: Per-motion ablation on paired-motion data augmentation, shown for both contact bodies and impulse. Solid arrows show our full method (Ours); dotted arrows show the no-augmentation baseline (No-aug). Both arrows go from T ✔ near to T ✘ near. Without augmentation, the arrows are short or even reverse direction, indicating the policy fails to follow the contact label [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] 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

50 extracted references · 50 canonical work pages · 14 internal anchors

  1. [1]

    Q. Liao, T. E. Truong, X. Huang, Y . Gao, G. Tevet, K. Sreenath, and C. K. Liu. Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion.arXiv preprint arXiv:2508.08241, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    M. Ji, X. Peng, F. Liu, J. Li, G. Yang, X. Cheng, and X. Wang. Exbody2: Advanced expressive humanoid whole-body control.arXiv preprint arXiv:2412.13196, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi. Om- nih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. InConference on Robot Learning (CoRL), 2024

    work page 2024

  4. [4]

    Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, and C. Finn. Humanplus: Humanoid shadowing and imitation from humans. InConference on Robot Learning (CoRL), 2024

    work page 2024

  5. [5]

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbabu, C. Pan, Z. Yi, G. Qu, K. Kitani, J. Hodgins, L. Fan, Y . Zhu, C. Liu, and G. Shi. Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills. InRobotics: Science and Systems (RSS), 2025

    work page 2025

  6. [6]

    Tessler, Y

    C. Tessler, Y . Guo, O. Nabati, G. Chechik, and X. B. Peng. Maskedmimic: Unified physics- based character control through masked motion inpainting.ACM Transactions on Graphics (SIGGRAPH Asia), 2024

    work page 2024

  7. [7]

    T. He, W. Xiao, T. Lin, Z. Luo, Z. Xu, Z. Jiang, J. Kautz, C. Liu, G. Shi, X. Wang, L. Fan, and Y . Zhu. Hover: Versatile neural whole-body controller for humanoid robots.arXiv preprint arXiv:2410.21229, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu. Perpetual humanoid control for real-time simulated avatars. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    work page 2023

  9. [9]

    Cheng, Y

    X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang. Expressive whole-body control for humanoid robots. InRobotics: Science and Systems (RSS), 2024

    work page 2024

  10. [10]

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi. Learning human-to-humanoid real-time whole-body teleoperation. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024

    work page 2024

  11. [11]

    Z. Chen, M. Ji, X. Cheng, X. Peng, X. B. Peng, and X. Wang. Gmt: General motion tracking for humanoid whole-body control.arXiv preprint arXiv:2506.14770, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    W. Xie, J. Han, J. Zheng, H. Li, X. Liu, J. Shi, W. Zhang, C. Bai, and X. Li. Kung- fubot: Physics-based humanoid whole-body control for learning highly-dynamic skills.arXiv preprint arXiv:2506.12851, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Y . Ze, Z. Chen, J. P. Ara´ujo, Z.-a. Cao, X. B. Peng, J. Wu, and C. K. Liu. Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    J. Li, X. Cheng, T. Huang, S. Yang, R.-Z. Qiu, and X. Wang. Amo: Adaptive motion optimiza- tion for hyper-dexterous humanoid whole-body control.arXiv preprint arXiv:2505.03738, 2025. 9

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Visual Imitation Enables Contextual Humanoid Control

    A. Allshire, H. Choi, J. Zhang, D. McAllister, A. Zhang, C. M. Kim, T. Darrell, P. Abbeel, J. Malik, and A. Kanazawa. Visual imitation enables contextual humanoid control.arXiv preprint arXiv:2505.03729, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  16. [16]

    Q. Ben, F. Jia, J. Zeng, J. Dong, D. Lin, and J. Pang. Homie: Humanoid loco-manipulation with isomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Y . Ze, Z. Chen, W. Wang, T. Chen, X. He, Y . Yuan, X. B. Peng, and J. Wu. Gener- alizable humanoid manipulation with 3d diffusion policies. In2025 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS), pages 2873–2880, 2025. doi: 10.1109/IROS60139.2025.11246340

    work page doi:10.1109/iros60139.2025.11246340 2025

  18. [18]

    X. He, S. Xu, X. Li, R. Dong, L. Bian, Y .-X. Wang, and L.-Y . Gui. Ultra: Unified mul- timodal control for autonomous humanoid whole-body loco-manipulation.arXiv preprint arXiv:2603.03279, 2026

    work page arXiv 2026

  19. [19]

    R. Dong, Z. Li, X. He, and S. Gupta. Learning humanoid end-effector control for open- vocabulary visual loco-manipulation.arXiv preprint arXiv:2602.16705, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  20. [20]

    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

    work page arXiv 2025

  21. [21]

    S. Yin, Y . Ze, H.-X. Yu, C. K. Liu, and J. Wu. Visualmimic: Visual humanoid loco- manipulation via motion tracking and generation.arXiv preprint arXiv:2509.20322, 2025

    work page arXiv 2025

  22. [22]

    F. Liu, Z. Gu, Y . Cai, Z. Zhou, H. Jung, J. Jang, S. Zhao, S. Ha, Y . Chen, D. Xu, and Y . Zhao. Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco- manipulation.arXiv preprint arXiv:2409.20514, 2024

    work page arXiv 2024

  23. [23]

    S. Zhao, Y . Ze, Y . Wang, C. K. Liu, P. Abbeel, G. Shi, and R. Duan. Resmimic: From gen- eral motion tracking to humanoid whole-body loco-manipulation via residual learning.arXiv preprint arXiv:2510.05070, 2025

    work page arXiv 2025

  24. [24]

    Mahmood, N

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black. Amass: Archive of motion capture as surface shapes. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019

    work page 2019

  25. [25]

    Taheri, N

    O. Taheri, N. Ghorbani, M. J. Black, and D. Tzionas. Grab: A dataset of whole-body human grasping of objects. InEuropean Conference on Computer Vision (ECCV), 2020

    work page 2020

  26. [26]

    B. L. Bhatnagar, X. Xie, I. Petrov, C. Sminchisescu, C. Theobalt, and G. Pons-Moll. Be- have: Dataset and method for tracking human object interactions. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  27. [27]

    J. Li, J. Wu, and C. K. Liu. Object motion guided human motion synthesis.ACM Transactions on Graphics (SIGGRAPH Asia), 2023

    work page 2023

  28. [28]

    Lu, C.-H

    J. Lu, C.-H. P. Huang, U. Bhattacharya, Q. Huang, and Y . Zhou. Humoto: A 4d dataset of mocap human object interactions. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025

    work page 2025

  29. [29]

    L. Yang, X. Huang, Z. Wu, A. Kanazawa, P. Abbeel, C. Sferrazza, C. K. Liu, R. Duan, and G. Shi. Omniretarget: Interaction-preserving data generation for humanoid whole-body loco- manipulation and scene interaction.arXiv preprint arXiv:2509.26633, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    J. P. Araujo, Y . Ze, P. Xu, J. Wu, and C. K. Liu. Retargeting matters: General motion retargeting for humanoid motion tracking.IEEE International Conference on Robotics and Automation (ICRA), 2026. 10

    work page 2026

  31. [31]

    Mordatch, E

    I. Mordatch, E. Todorov, and Z. Popovi ´c. Discovery of complex behaviors through contact- invariant optimization.ACM Transactions on Graphics (SIGGRAPH), 31(4), 2012

    work page 2012

  32. [32]

    M. Posa, C. Cantu, and R. Tedrake. A direct method for trajectory optimization of rigid bodies through contact.The International Journal of Robotics Research (IJRR), 33(1):69–81, 2014

    work page 2014

  33. [33]

    Tassa, T

    Y . Tassa, T. Erez, and E. Todorov. Synthesis and stabilization of complex behaviors through online trajectory optimization. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4906–4913, 2012

    work page 2012

  34. [34]

    X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills.ACM Transactions on Graphics (SIG- GRAPH), 2018

    work page 2018

  35. [35]

    Starke, H

    S. Starke, H. Zhang, T. Komura, and J. Saito. Neural state machine for character-scene inter- actions.ACM Transactions on Graphics (SIGGRAPH Asia), 2019

    work page 2019

  36. [36]

    Hassan, D

    M. Hassan, D. Ceylan, R. Villegas, J. Saito, J. Yang, Y . Zhou, and M. J. Black. Stochastic scene-aware motion prediction. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    work page 2021

  37. [37]

    S. Xu, Z. Li, Y .-X. Wang, and L.-Y . Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    work page 2023

  38. [38]

    X. Peng, Y . Xie, Z. Wu, V . Jampani, D. Sun, and H. Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models.arXiv preprint arXiv:2312.06553, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  39. [39]

    J. Li, A. Clegg, R. Mottaghi, J. Wu, X. Puig, and C. K. Liu. Controllable human-object interaction synthesis. InEuropean Conference on Computer Vision (ECCV), 2024

    work page 2024

  40. [40]

    X. B. Peng, A. Kanazawa, J. Malik, P. Abbeel, and S. Levine. Sfv: Reinforcement learning of physical skills from videos.ACM Transactions on Graphics (TOG), 37(6), 2018

    work page 2018

  41. [41]

    S. Xu, S. Schulter, M. Ziyadi, X. He, X. Fei, Y .-X. Wang, and L.-Y . Gui. Interprior: Scaling generative control for physics-based human-object interactions.IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    work page 2026

  42. [42]

    L. Pan, Z. Yang, Z. Dou, W. Wang, B. Huang, B. Dai, T. Komura, and J. Wang. Tokenhsi: Unified synthesis of physical human-scene interactions through task tokenization.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  43. [43]

    Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning

    M. Mittal, P. Roth, J. Tigue, A. Richard, et al. Isaac lab: A gpu-accelerated simulation frame- work for multi-modal robot learning.arXiv preprint arXiv:2511.04831, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    Proximal Policy Optimization Algorithms

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017. 11 APPENDIX We provide additional details and results that supplement CONTACTMIMIC:

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  45. [45]

    § A describes the organization of the overview video and the project page

    work page

  46. [46]

    § B provides additional implementation details, including PPO hyperparameters, policy obser- vations and architecture, and domain randomization

    work page

  47. [47]

    § C lists the full set of reward terms with their weights and shaping parameters

    work page

  48. [48]

    § D details our paired-motion generation pipeline: contact-label extraction, augmentation pa- rameters, and the per-motion training configuration

    work page

  49. [49]

    § E describes our real-world deployment on the Unitree G1

    work page

  50. [50]

    no- contact

    § F reports additional quantitative results, including the full per-motion contact controllability table and the per-motion data augmentation ablation. A Overview Video and Project Page We provide an overview video and the raw recordings of every real-world trial reported in Table 2 on our project page. The videos and page include: •Method overview: the f...

    work page 1913