Computer Vision and Pattern Recognition · arXiv:2611.11517 v1 · submitted 8 Nov 2026 · 76 references
InfCoiL: Coordinated Planner-Controller Learning for Closed-Loop Physics-Based Human-Object Interaction
The paper says
“Text- and goal-conditioned human-object interaction (HOI) provides an intuitive interface for specifying high-level intents in physics-based character animation.”
more of the abstract
Mapping such intents to physically plausible, contact-rich behaviors naturally calls for a generative planner and a physics-based controller, yet existing planner-controller systems either execute plans open-loop or close the loop while keeping the planner fixed, so execution feedback never reshapes what is planned. We present InfCoiL, a closed-loop framework for physics-based HOI from coordinated planner-controller learning. InfCoiL couples a morphology-aware flow-matching planner with a unified multi-morphology controller through a shared morphology interface in an autoregressive plan-and-execute loop, and alternately fine-tunes both components with reinforcement learning from the same physical rollouts. To optimize the flow-matching planner online, we introduce Semantic Flow Policy Optimization (SFPO), which stochasticizes selected denoising transitions with channel-wise semantic covariance over heterogeneous HOI features and assigns group-relative, sequence-level rewards from complete closed-loop executions. On held-out OMOMO sequences, InfCoiL outperforms open-loop hierarchical baselines in goal-reaching task success and physical tracking accuracy. Project page: https://infcoil.github.io/
from the abstract
Pith says
Machine review · deepseek-v4-flash · 10 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Solid and well-executed: the closed-loop planner-controller framework with SFPO is a genuine advance, but the lack of error bars and the hand-tuned noise scales are the main soft spots. This line is wrong
2 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 76 references say.
The paper says, at §2.2.1, Table 7, Appendix F.2
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.89):
“Since low precision has been linked to loss spikes [75] and to training-inference mismatch in RL fine-tuning [76], the numerical precision and the hand-set noise scales κ deserve a joint study.”
Pith says: Objection 1 of 2 · major
The SFPO planner update depends on hand-chosen channel-wise noise scales κ (Table 7). The paper tests only the global multiplier λ_n (Table 2) and does not vary the per-channel κ values. The authors acknowledge in Appendix F.2 that 'the hand-set noise scales κ deserve a joint study.'
more of objection 1
This is load-bearing because the exploration strategy of SFPO is defined by these scales; if performance degrades significantly under plausible perturbations of κ (e.g., ±50% on one channel group), the claim that SFPO is a robust and general planner update would be weakened. I consider this a major gap that should be addressed, at minimum by a sensitivity analysis on a subset of channel groups or by providing guidance for setting κ in new tasks.
The paper says, at §3.2, Tables 1 and 2, §3.1
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The main evaluation reports only point estimates (SR@d, goal errors, tracking errors) with no standard deviations, confidence intervals, or results across multiple random seeds. Although each test sequence is evaluated three times and the metrics are averaged, the paper does not report the variance across those runs or across training seeds.
more of objection 2
This makes it difficult to assess the statistical significance of the reported improvements, especially for comparisons where differences are moderate (e.g., 4.17% SR between w/o FT P and C→P, or 1.84% between SFPO at λ_n=0.1 and λ_n=0.2). The authors should provide error bars or at least report the variance over multiple independent evaluations.
5 minor objections
The paper says, at Abstract and §1
Pith says: minor objection 1 of 5
The phrase 'closes three coupled loops (interface, execution, learning)' is evocative but could be clarified early in the body. The distinction between the interface loop and the execution loop is not immediately clear from the description.
The paper says, at §2.1.2, Eq. (4)
Pith says: minor objection 2 of 5
The equation v_θ = (x_τ - f_θ) / τ is unambiguous, but adding parentheses in the text ('(x_τ - f_θ) / τ') would prevent any misreading.
show 3 more
The paper says, at Table 1
Pith says: minor objection 3 of 5
The column headers 'Pelvis Goal↓' and 'Object Goal↓' should explicitly state that the values are in centimeters (e.g., 'Pelvis Goal (cm)↓'), as done in the surrounding text.
The paper says, at Appendix E.1
Pith says: minor objection 4 of 5
The symbol ˜τ_k is introduced in the sentence 'with ˜τ_k = τ_{k+1} at the endpoint τ_k=1 and ˜τ_k = τ_k otherwise,' but the expression √(τ_k / max(1-˜τ_k, 1e-6)) appears before this definition. Move the definition earlier or restructure the paragraph.
The paper says, at References
Pith says: minor objection 5 of 5
Several references (e.g., [3], [5], [16], [17]) are to arXiv preprints without peer-reviewed publication status noted. While acceptable, the authors may wish to indicate which versions are used, or update citations where conference versions exist.
What would settle it
Test sensitivity of SFPO to the per-channel κ by a factorial perturbation: vary each of the six κ groups independently by factors of ×0.5, ×1.0, ×2.0 (full factorial would be 3^6=729, too many; instead randomly sample 12 combinations uniform over these three factors for each group). Use the best λ_n=0.1 and run the full InfCoiL (C→P→C) pipeline on OMOMO test. Record SR@0.3 for each combination. If the minimum SR@0.3 across the 12 runs drops below 71.19% (i.e., >10 point drop from 81.19%), the method is sensitive to κ tuning. If all 12 runs stay within 5 points of the reported value, robustness is supported.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 10 Nov 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Here's my take on InfCoiL. The paper delivers what it promises: a clean closed-loop architecture that jointly fine-tunes a flow-matching planner and a physics controller via alternating RL. The key technical novelty is SFPO, which turns a subset of denoising steps into Gaussian policy steps with channel-wise noise scales matched to heterogeneous HOI tokens. That's a sensible design choice for making flow-matching planners trainable with on-policy rewards, and the ablations show it works where isotropic baselines fail. The empirical story is strong: 81% SR vs 24–52% for open-loop cascades on the OMOMO test set, with component-wise ablations that isolate the effect of replanning, controller adaptation, and planner learning. The matched-budget comparison (Appendix B.3) is particularly convincing—the gain comes from co-adaptation, not more controller iterations.
Two things keep this from being a clear accept. First, no error bars on the main results. The paper states 'each test sequence is evaluated three times and metrics are averaged' but never reports variance across those three runs or across random seeds. Given the stochasticity in RL training and the small number of test sequences, some readers will want to know whether the lead over baselines is statistically meaningful. Second, the SFPO noise scales (Table 7) are hand-picked and tested only via the global multiplier λn; the per-channel κ are not varied. The authors themselves flag this in Appendix F.2 as deserving 'a joint study.' That said, the paper does report three λn values (0.05, 0.1, 0.2) with consistent improvements, and the main conclusion does not rely on a fragile calibration—λn=0.05 still beats the starting point. The sensitivity concern is real but not load-bearing for the central claim.
Overall, this is a serious piece of work by thoughtful researchers. The contribution is clear, the experiments are thorough (within the reported scope), and the limitations are stated honestly. I would accept this for peer review with a request for error bars and a brief sensitivity analysis on κ or a simple adaptive scheme.
Who is this for: researchers working on physics-based character animation, hierarchical RL for human-object interaction, or diffusion/flow-based motion planning. The paper deserves a serious referee.
The strongest objection
The SFPO planner update introduces exploration through a diagonal covariance with hand-set, channel-group-dependent scales κ (Table 7: body 0.35, hand 0.175, rotation 0.07, object translation 0.7, object rotation 0.07, contact 0.175). The paper tests only three values of the global multiplier λ_n, not the per-channel κ. The authors themselves flag in Appendix F.2 that 'the hand-set noise scales κ deserve a joint study.' This is load-bearing because the entire SFPO exploration strategy rests on these scales: if a different κ for, say, object translation leads to a large drop in success rate, then the coordinated learning pipeline becomes brittle and its applicability to new tasks or morphologies would require extensive re-tuning. While Table 2 shows that at λ_n=0.1 the method achieves 77.24% SR@0.3, it does not address whether the specific κ vector is critical or whether a wide range of κ would still work. The central claim—that coordinated closed-loop learning via SFPO yields substantial gains—would be undermined if the planner update only functions near the calibrated κ. Moreover, the log-probability ratio in Eq. (24) averages over elements, so mis-scaled κ could distort the importance ratios. The reader's weakest assumption already targets hyperparameter sensitivity; this concretizes it.
The reader identified the hand-chosen SFPO noise scales as the weakest assumption. My concern sharpens this: the κ values themselves are untested, not just the global λ_n. The paper's own admission in Appendix F.2 confirms this is an unresolved point. If the method is sensitive to κ, the central claim of a general closed-loop coordination framework would be weakened, requiring significant re-tuning for each new setting. The proposed concrete test directly probes this vulnerability by varying κ within a plausible range (halving/doubling) and checking whether performance collapses. Since the reader already recommended CONDITIONAL acceptance based on hyperparameter sensitivity, and my concern reinforces rather than reverses that assessment, the verdict should remain UNCHANGED. The test would provide the missing evidence; if robust, the condition would be satisfied, potentially moving to ACCEPT. The reader and I agree on the weakest assumption, so agreement_with_reader is 'agree'.
The referee's report
If the results hold, this work makes an important step in hierarchical physics-based HOI by demonstrating that closed-loop co-adaptation of a generative planner and a physics-based controller can significantly improve task success over open-loop cascades. The SFPO method provides a practical way to apply online RL to flow-matching planners without requiring a separate score model. The thorough ablations (closed-loop replanning alone, controller-only, planner-only, alternating phases) clearly show that coordinated learning raises the performance ceiling beyond what either component alone can achieve. This could influence future research on closed-loop planning for contact-rich character animation and embodied agents.
Pith's reading of the paper (a machine's summary, not the authors')
This paper proposes InfCoiL, a closed-loop framework that couples a morphology-aware flow-matching planner with a unified multi-morphology controller for physics-based human-object interaction. The key innovation is coordinated learning: the planner and controller are alternately fine-tuned using reinforcement learning from the same physical rollouts, so execution feedback reshapes both what is planned and how it is tracked. On held-out OMOMO sequences, InfCoiL raises task success rate from 24–52% (open-loop baselines) to 81%, while reducing pelvis and object goal errors by more than half. Ablations show that closed-loop replanning corrects recoverable drift, controller adaptation enlarges the executable region, and planner learning (via the proposed SFPO) further raises the success ceiling after the controller has expanded.
What carries the argument. Semantic Flow Policy Optimization (SFPO) converts the final denoising step of a flow-matching planner into a Gaussian policy step with channel-wise semantic covariance, enabling tractable log-probability computation and structured exploration across heterogeneous HOI tokens (body, object, contact). SFPO then assigns group-relative advantages based on complete closed-loop execution scores to update the planner, while the controller is updated concurrently with PPO using the same physical rollouts.
Core claim. The central claim is that coordinated closed-loop fine-tuning of both a flow-matching planner and a physics-based controller via alternating PPO (controller) and SFPO (planner) significantly outperforms open-loop hierarchies or single-module fine-tuning. The paper demonstrates this on the OMOMO benchmark, reporting a success rate at 0.3 m threshold of 81.19% for InfCoiL (C→P→C) compared to 49.23% for the open-loop variant and 24.62% for the InterMimic student baseline. The improvement is attributed to the closed-loop loop and the alternating learning strategy that aligns the planner's generative distribution with the controller's executable behavior distribution.
The premise it rests on. SFPO relies on hand-chosen channel-wise noise scales κ and a global noise level λ_n that were tuned on the OMOMO validation set; the paper provides no theoretical justification for these scales nor a sensitivity analysis across different datasets or tasks, so the planner update could fail on new objects or morphologies without extensive retuning.
If the paper is right (Pith's inferences, not the authors')
Physics-based character animation can achieve substantially higher task success by closing the loop between a generative planner and a tracking controller, rather than executing a one-shot plan.
Flow-matching generative models can be fine-tuned online with RL from sparse execution feedback, using a stochastic denoising path that respects the heterogeneous sensitivity of different motion channels.
Alternating planner and controller fine-tuning is more effective than training either component alone, because controller adaptation first expands the set of executable references, after which planner learning can exploit that larger region.
The framework may generalize to other hierarchical control problems where a high-level generator must adapt to a low-level executor's capabilities.
SFPO's hand-chosen channel-wise noise scales κ and global level λ_n were tuned on the OMOMO validation set; the method's sensitivity to these hyperparameters on new objects or morphologies is unknown and could be a brittle point.
The closed-loop scheme assumes structured object states and goal specifications; extending it to visual perception inputs (e.g., from cameras) would require an additional perception module, but the coordinated learning paradigm itself could carry over.
If the noise scales prove robust across domains, the SFPO mechanism could be applied to other flow-matching or diffusion planners in robotics, where different sensor modalities also have heterogeneous error tolerances.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Text- and goal-conditioned human-object interaction (HOI) provides an intuitive interface for specifying high-level intents in physics-based character animation. Mapping such intents to physically plausible, contact-rich behaviors naturally calls for a generative planner and a physics-based controller, yet existing planner-controller systems either execute plans open-loop or close the loop while keeping the planner fixed, so execution feedback never reshapes what is planned. We present InfCoiL, a closed-loop framework for physics-based HOI from coordinated planner-controller learning. InfCoiL couples a morphology-aware flow-matching planner with a unified multi-morphology controller through a shared morphology interface in an autoregressive plan-and-execute loop, and alternately fine-tunes both components with reinforcement learning from the same physical rollouts. To optimize the flow-matching planner online, we introduce Semantic Flow Policy Optimization (SFPO), which stochasticizes selected denoising transitions with channel-wise semantic covariance over heterogeneous HOI features and assigns group-relative, sequence-level rewards from complete closed-loop executions. On held-out OMOMO sequences, InfCoiL outperforms open-loop hierarchical baselines in goal-reaching task success and physical tracking accuracy. Project page: https://infcoil.github.io/
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-10
The authors' main claim, as the classifier picked it out 0.99
“On held-out OMOMO sequences, InfCoiL outperforms open-loop hierarchical baselines in goal-reaching task success and physical tracking accuracy.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.11 · states limitation 0.10 · human participants 0.08 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.06 · compares with prior results 0.95 · mentions code or data availability 0.30
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Text- and goal-conditioned human-object interaction (HOI) provides an intuitive interface for specifying high-level intents in physics-based character animation. | background 1.00 | 0.06 | 0.02 | asserted 1.00 |
| Mapping such intents to physically plausible, contact-rich behaviors naturally calls for a generative planner and a physics-based controller, yet existing planner-controller systems either execute plans open-loop or close the loop while keeping the planner fixed, so execution feedback never reshapes what is planned. | gap 0.97 | 0.06 | 0.06 | asserted 0.94 |
| We present InfCoiL, a closed-loop framework for physics-based HOI from coordinated planner-controller learning. | contribution 1.00 | 0.90 | 0.23 | asserted 1.00 |
| InfCoiL couples a morphology-aware flow-matching planner with a unified multi-morphology controller through a shared morphology interface in an autoregressive plan-and-execute loop, and alternately fine-tunes both components with reinforcement learning from the same physical rollouts. | method 0.93 | 0.94 | 0.11 | asserted 1.00 |
| To optimize the flow-matching planner online, we introduce Semantic Flow Policy Optimization (SFPO), which stochasticizes selected denoising transitions with channel-wise semantic covariance over heterogeneous HOI features and assigns group-relative, sequence-level rewards from complete closed-loop executions. | method 0.86 | 0.95 | 0.05 | asserted 0.99 |
| On held-out OMOMO sequences, InfCoiL outperforms open-loop hierarchical baselines in goal-reaching task success and physical tracking accuracy. | result 1.00 | 0.96 | 0.59 | asserted 1.00 |
| Project page: https://infcoil.github.io/ | future work 0.55 | 0.12 | 0.02 | unclear 0.97 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 0.69
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-10: 1 request, 5,716 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (4)
Figures are the authors', reproduced from the paper.
References (76)
- [1] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928–14940, 2023 on Pith
- [2] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. InEuropean Conference on Computer Vision, pages 54–72. Springer, 2024. 10 on Pith
- [3] Ziyin Wang, Sirui Xu, Chuan Guo, Bing Zhou, Jiangshan Gong, Jian Wang, Yu-Xiong Wang, and Liang- Yan Gui. Unleashing guidance without classifiers for human-object interaction animation.arXiv preprint arXiv:2603.25734, 2026 on PitharXiv
- [4] Sirui Xu, Hung Yu Ling, Yu-Xiong Wang, and Liang-Yan Gui. Intermimic: Towards universal whole-body control for physics-based human-object interactions. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12266–12277, 2025 on Pith
- [5] Sirui Xu, Samuel Schulter, Morteza Ziyadi, Xialin He, Xiaohan Fei, Yu-Xiong Wang, and Liangyan Gui. Interprior: Scaling generative control for physics-based human-object interactions.arXiv preprint arXiv:2602.06035, 2026 on PitharXiv
- [6] Yuhang Lin, Yijia Xie, Jiahong Xie, Yuehao Huang, Ruoyu Wang, Jiajun Lv, Yukai Ma, and Xingxing Zuo. Simgenhoi: Physically realistic whole-body humanoid-object interaction via generative modeling and reinforcement learning.arXiv preprint arXiv:2508.14120, 2025 on PithPith's reviewarXiv
- [7] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control.arXiv preprint arXiv:2410.03441, 2024 on PithPith's reviewarXiv
- [8] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis.ACM Transactions on Graphics (TOG), 42(6):1–11, 2023 on Pith
- [9] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio... on Pith
- [10] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. InProceedings of the IEEE/CVF international conference on computer vision, pages 4332–4341, 2019 on Pith
- [11] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022 on PithPith's reviewarXiv
- [12] Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise.arXiv preprint arXiv:2511.13720, 2025 on PithPith's reviewarXiv
- [13] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills.ACM Transactions On Graphics (TOG), 37(4):1–14, 2018 on Pith
- [14] Patrick Yin, Tyler Westenbroek, Zhengyu Zhang, Joshua Tran, Ignacio Dagnino, Eeshani Shilamkar, Numfor Mbiziwo-Tiapo, Simran Bagaria, Xinlei Liu, Galen Mullins, et al. Emergent dexterity via diverse resets and large-scale reinforcement learning.arXiv preprint arXiv:2603.15789, 2026 on PitharXiv
- [15] Tonghe Zhang, Chao Yu, Sichang Su, and Yu Wang. Reinflow: Fine-tuning flow matching policy with online reinforcement learning.arXiv preprint arXiv:2505.22094, 2025 on PitharXiv
- [16] Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470, 2025 on PithPith's reviewarXiv
- [17] David McAllister, Songwei Ge, Brent Yi, Chung Min Kim, Ethan Weber, Hongsuk Choi, Haiwen Feng, and Angjoo Kanazawa. Flow matching policy gradients.arXiv preprint arXiv:2507.21053, 2025 on PithPith's reviewarXiv
- [18] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024 on PithPith's reviewarXiv
- [19] Brent Yi, Hongsuk Choi, Himanshu Gaurav Singh, Xiaoyu Huang, Takara E Truong, Carmelo Sferrazza, Yi Ma, Rocky Duan, Pieter Abbeel, Guanya Shi, et al. Flow policy gradients for robot control.arXiv preprint arXiv:2602.02481, 2026 on PithPith's reviewarXiv
- [20] Christian Diller and Angela Dai. Cg-hoi: Contact-guided 3d human-object interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19888– 19901, 2024. 11 on Pith
show 56 more
- [21] Peishan Cong, Ziyi Wang, Yuexin Ma, and Xiangyu Yue. Semgeomo: Dynamic contextual human motion generation with semantic and geometric guidance. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 17561–17570, 2025 on Pith
- [22] Mengqing Xue, Yifei Liu, Ling Guo, Shaoli Huang, and Changxing Ding. Guiding human-object interactions with rich geometry and relations. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22714–22723, 2025 on Pith
- [23] Ling-An Zeng, Guohong Huang, Yi-Lin Wei, Shengbo Gu, Yu-Ming Tang, Jingke Meng, and Wei-Shi Zheng. Chainhoi: Joint-based kinematic chain modeling for human-object interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page... on Pith
- [24] Shujia Li, Haiyu Zhang, Xinyuan Chen, Yaohui Wang, and Yutong Ban. Genhoi: Generalizing text-driven 4d human-object interaction synthesis for unseen objects.arXiv preprint arXiv:2506.15483, 2025 on PithPith's reviewarXiv
- [25] Lin Wu, Zhixiang Chen, and Jianglin Lan. Hoi-dyn: Learning interaction dynamics for human-object motion diffusion.arXiv preprint arXiv:2507.01737, 2025 on PitharXiv
- [26] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real- time controllable motion generation via latent consistency model. InEuropean Conference on Computer Vision, pages 390–408. Springer, 2024 on Pith
- [27] Lei Jiang, Ye Wei, and Hao Ni. Motionpcm: Real-time motion synthesis with phased consistency model. arXiv preprint arXiv:2501.19083, 2025 on PithPith's reviewarXiv
- [28] Yude Zou, Junji Gong, Xing Gao, Zixuan Li, Tianxing Chen, and Guanjie Zheng. Infbagel: Human- object-scene interaction generation with dynamic perception and iterative refinement.arXiv preprint arXiv:2604.04843, 2026 on PithPith's reviewarXiv
- [29] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks.ACM Transactions on Graphics (TOG), 39(4):39–1, 2020 on Pith
- [30] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning.Acm transactions on graphics (tog), 37(4):1–14, 2018 on Pith
- [31] Xinyu Xu, Yizheng Zhang, Yong-Lu Li, Lei Han, and Cewu Lu. Humanvla: Towards vision-language directed object rearrangement by physical humanoid.Advances in Neural Information Processing Systems, 37:18633–18659, 2024 on Pith
- [32] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. Physhoi: Physics-based imitation of dynamic human-object interaction.arXiv preprint arXiv:2312.04393, 2023 on PithPith's reviewarXiv
- [33] Yinhuai Wang, Qihan Zhao, Runyi Yu, Hok Wai Tsui, Ailing Zeng, Jing Lin, Zhengyi Luo, Jiwen Yu, Xiu Li, Qifeng Chen, et al. Skillmimic: Learning basketball interaction skills from demonstrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit... on Pith
- [34] Runyi Yu, Yinhuai Wang, Qihan Zhao, Hok Wai Tsui, Jingbo Wang, Ping Tan, and Qifeng Chen. Skillmimic- v2: Learning robust and generalizable interaction skills from sparse and noisy demonstrations. InPro- ceedings of the Special Interest Group on Computer Graphics and Interacti... on Pith
- [35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag... on Pith
- [36] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895–10904, 2023 on Pith
- [37] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting.ACM Transactions On Graphics (TOG), 43(6):1–21, 2024 on Pith
- [38] Yifei Yao, Chengyuan Luo, Jiaheng Du, Wentao He, and Jun-Guo Lu. Gbc: Generalized behavior-cloning framework for whole-body humanoid imitation.arXiv preprint arXiv:2508.09960, 2025. 12 on PithPith's reviewarXiv
- [39] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters.ACM Transactions On Graphics (TOG), 41(4):1–17, 2022 on Pith
- [40] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control.ACM Transactions on Graphics (ToG), 40(4):1–20, 2021 on Pith
- [41] Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations for physics-based control.arXiv preprint arXiv:2310.04582, 2023 on PithPith's reviewarXiv
- [42] Heyuan Yao, Zhenhua Song, Baoquan Chen, and Libin Liu. Controlvae: Model-based learning of generative controllers for physics-based characters.ACM Transactions on Graphics (TOG), 41(6):1–16, 2022 on Pith
- [43] Heyuan Yao, Zhenhua Song, Yuyang Zhou, Tenglong Ao, Baoquan Chen, and Libin Liu. Moconvq: Unified physics-based motion control via scalable discrete representations.ACM Transactions on Graphics (TOG), 43(4):1–21, 2024 on Pith
- [44] Yan Wu, Korrawe Karunratanakul, Zhengyi Luo, and Siyu Tang. Uniphys: Unified planner and controller with diffusion for flexible physics-based character control. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13214–13224, 2025 on Pith
- [45] Xiaoyu Huang, Takara Truong, Yunbo Zhang, Fangzhou Yu, Jean Pierre Sleiman, Jessica Hodgins, Koushil Sreenath, and Farbod Farshidian. Diffuse-cloc: Guided diffusion for physics-based character look-ahead control.ACM Transactions on Graphics (TOG), 44(4):1–12, 2025 on Pith
- [46] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. Superpadl: Scaling language-directed physics-based control with progressive supervised distillation. InACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024 on Pith
- [47] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Condi- tional adversarial latent models for directable virtual characters. InACM SIGGRAPH 2023 conference proceedings, pages 1–9, 2023 on Pith
- [48] Takara Everest Truong, Michael Piseno, Zhaoming Xie, and Karen Liu. Pdp: Physics-based character animation via diffusion policy. InSIGGRAPH Asia 2024 Conference Papers, pages 1–10, 2024 on Pith
- [49] Nicklas Hansen, Jyothir SV , Vlad Sobal, Yann LeCun, Xiaolong Wang, and Hao Su. Hierarchical world models as visual whole-body humanoid controllers.arXiv preprint arXiv:2405.18418, 2024 on PithPith's reviewarXiv
- [50] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag... on Pith
- [51] Zhixuan Liang, Yao Mu, Mingyu Ding, Fei Ni, Masayoshi Tomizuka, and Ping Luo. Adaptdiffuser: Diffusion models as adaptive self-evolving planners.arXiv preprint arXiv:2302.01877, 2023 on PithPith's reviewarXiv
- [52] Zixuan Wu, Sean Ye, Manisha Natarajan, and Matthew C Gombolay. Diffusion-reinforcement learning hierarchical motion planning in adversarial multi-agent games.arXiv e-prints, pages arXiv–2403, 2024 on Pith
- [53] Liangliang Chen, Yutian Lei, Shiyu Jin, Ying Zhang, and Liangjun Zhang. Rlingua: Improving reinforce- ment learning sample efficiency in robotic manipulations with large language models.IEEE Robotics and Automation Letters, 9(7):6075–6082, 2024 on Pith
- [54] Yifei Yao, Wentao He, Chenyu Gu, Jiaheng Du, Fuwei Tan, Zhen Zhu, and Jun-Guo Lu. Anybipe: An automated end-to-end framework for training and deploying bipedal robots powered by large language models. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (... on Pith
- [55] Allen Ren, Justin Lidard, Lars Ankile, Anthony Simeonov, Pulkit Agrawal, Anirudha Majumdar, Benjamin Burchfiel, Hongkai Dai, and Max Simchowitz. Diffusion policy policy optimization. InInternational Conference on Learning Representations, volume 2025, pages 77288–77329, 2025 on Pith
- [56] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. InEuropean conference on computer vision, pages 581–600. Springer, 2020. 13 on Pith
- [57] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 159... on Pith
- [58] Jiaxin Lu, Chun-Hao Paul Huang, Uttaran Bhattacharya, Qixing Huang, and Yi Zhou. Humoto: A 4d dataset of mocap human object interactions. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10886–10897, 2025 on Pith
- [59] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. Coohoi: Learning cooperative human-object interaction with manipulated object dynamics.Advances in Neural Information Processing Systems, 37:79741–79763, 2024 on Pith
- [60] Liang Pan, Zeshi Yang, Zhiyang Dou, Wenjia Wang, Buzhen Huang, Bo Dai, Taku Komura, and Jingbo Wang. Tokenhsi: Unified synthesis of physical human-scene interactions through task tokenization. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5379... on Pith
- [61] Boyu Mi, Mengchen Ma, Yifei Yao, Xing Gao, Junting Chen, Yangzi Li, Zihou Zhu, Guohao Li, Zhenfei Yin, Tai Wang, et al. Exploratory, communicative, and deployable: Vision-driven embodied agents for open-world mobile manipulation. InEuropean Conference on Computer Vision, pages... on Pith
- [62] Jisong Cai, Yao Mu, Ganlin Yang, Zhe Cao, Zhangzheng Tu, Xing Gao, Kailin Li, Xinyu Zhan, Lixin Yang, Yangkun Zhu, et al. Internw0: A foundational physical world model for efficient real-world interactions. arXiv preprint arXiv:2609.27656, 2026 on PithPith's reviewarXiv
- [63] Xingyu Miao, Zizun Li, Baole Fang, Kaiwen Song, Tenghui Wang, Hanxue Zhang, Yating Wang, Xudong Li, Yuping He, Xueyuan Wei, et al. Internw0-∆: A world action model bridging predictive dynamics and actions with 20k+ hours of open data.arXiv preprint arXiv:2609.31394, 2026 on PithPith's reviewarXiv
- [64] Yude Zou, Yifei Yao, Guanbo Jiao, Jianming Ma, Guanjie Zheng, and Xing Gao. PICA-WAM: Visual perturbation invariance via cross-appearance self-alignment in world action models.arXiv preprint arXiv:2610.26813, 2026 on PithPith's reviewarXiv
- [65] Yao Mu, Tianxing Chen, Shijia Peng, Zanxin Chen, Zeyu Gao, Yude Zou, Lunkai Lin, Zhiqiang Xie, and Ping Luo. Robotwin: Dual-arm robot benchmark with generative digital twins (early version). InEuropean Conference on Computer Vision, pages 264–273. Springer, 2024 on Pith
- [66] Yao Mu, Tianxing Chen, Zanxin Chen, Shijia Peng, Zhiqian Lan, Zeyu Gao, Zhixuan Liang, Qiaojun Yu, Yude Zou, Mingkun Xu, et al. Robotwin: Dual-arm robot benchmark with generative digital twins. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page... on Pith
- [67] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation.arXiv preprint ar... on PithPith's reviewarXiv
- [68] Ning Gao, Jinliang Zheng, Xing Gao, Haoxiang Ma, Hanqing Wang, Yukai Wang, Jiantong Chen, Zanxin Chen, Shujie Zhang, Mingda Jia, et al. Ebench: Elemental diagnosis of generalist mobile manipulation policies.arXiv preprint arXiv:2606.18239, 2026 on PithPith's reviewarXiv
- [69] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In2023 IEEE International conference on robotics and automation (ICRA), pages 9493–9500. IEEE, 2023 on Pith
- [70] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as I can, not as I say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691, 2022 on PithPith's reviewarXiv
- [71] Zhuo Lin, Sirui Xu, Liuyu Bian, Yu-Xiong Wang, and Liang-Yan Gui. InterEvolve: Test-time evolution of reward programs for humanoid loco-manipulation.arXiv preprint arXiv:2610.02196, 2026 on PithPith's reviewarXiv
- [72] Yifei Yao, Yude Zou, Guanbo Jiao, Xing Gao, Ning Gao, Yukai Wang, Jiantong Chen, Zihou Zhu, Xinyu Li, Ningyu Yan, et al. Profiling a frontier agent with EBench: From direct control to harness with policy. arXiv preprint arXiv:2610.35832, 2026 on PithPith's reviewarXiv
- [73] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed precision training. InInternational Conference on Learning Representations, 2018. 14 on Pith
- [74] Dhiraj Kalamkar, Dheevatsa Mudigere, Naveen Mellempudi, Dipankar Das, Kunal Banerjee, Sasikanth Avancha, Dharma Teja V ooturi, Nataraj Jammalamadaka, Jianyu Huang, Hector Yuen, et al. A study of BFLOAT16 for deep learning training.arXiv preprint arXiv:1905.12322, 2019 on PithPith's reviewarXiv
- [75] Hanqing Liu, Jianjun Cao, Yuanze Li, and Zijian Zhou. Grokking or glitching? How low-precision drives Slingshot loss spikes. InICML 2026 Workshop on High-Dimensional Learning Dynamics, 2026 on Pith
- [76] Penghui Qi, Zichen Liu, Xiangxin Zhou, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Defeating the training-inference mismatch via FP16.arXiv preprint arXiv:2510.26788, 2025. 15 NeurIPS Paper Checklist 1.Claims Question: Do the main claims made in the abstract and introducti... on PitharXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment