Pith. sign in

REVIEW 4 major objections 5 minor 35 references

IDAGC: Adaptive Generalized Human-Robot Collaboration via Human Intent Estimation and Multimodal Policy Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read IDAGC claims a single multimodal policy can estimate human intent, learn many tasks from few demonstrations, and switch between physical and remote collaboration modes on its own.

desk verdict A genuinely new integration of force-conditioned intent estimation with a multi-task policy, but the evaluation has a concrete data-count contradiction and too many point estimates to support the headline claims. read the letter →

arxiv 2507.04620 v1 pith:Z6E3HSPT submitted 2025-07-07 cs.RO

Haotian Liu , Yuchuang Tong , Guanchen Liu , Zhaojie Ju , Zhengtao Zhang This is my paper
classification cs.RO
keywords human-robotcollaborationintentestimationmultimodalpolicylearningconditionalvariationalautoencoderLIBERObenchmarkphysicalinteractionvision-language-actionmodelactionchunking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a single robot-learning framework, IDAGC, can handle both sides of human-robot collaboration at once: physical hand-in-hand interaction and remote cooperation guided by language and vision. By coupling a Transformer-based intent estimator that uses a conditional variational autoencoder (CVAE) with dedicated encoders for vision, language, force, and robot state, the framework learns multi-task policies from about 50 demonstrations per task and switches collaboration modes on its own. On the LIBERO benchmark, the authors report success rates of 91.1% across 90 tasks and 91.0% on 10 long-horizon tasks, ahead of the BAKU baseline. In a real-world obstacle-avoidance test, their intent encoder predicts human trajectories with an average error of 0.26 mm, compared with 1.94 mm for the nearest comparison method. If these numbers hold, the result would be a single policy that adapts its behavior to whether a human is steering physically or communicating remotely.

What carries the argument

The load-bearing object is the Transformer-based human intent encoder built on a conditional variational autoencoder (CVAE), which learns a latent variable $Z$ for the distribution of future trajectory $Y$ given past multimodal observations $X$. A past encoder produces a Gaussian prior $p(Z|X)$ from position, velocity, and force data, while a future encoder produces the posterior $q(Z|Y,X)$ during training; a future-past attention module aligns the two. The same latent representation is sampled at test time and fed into a Transformer decoder alongside FiLM-conditioned ResNet-18 image features, a MiniLM text encoder, robot-state MLP, and a causal-convolution force encoder, so that the decoded action chunks reflect the inferred human intention. The training loss combines the CVAE evidence lower bound with a behavior-cloning term, which is what lets one network do both intent prediction and policy learning.

What would settle it

Run the IDAGC policy on LIBERO-90 and LIBERO-10 with at least ten random seeds and report mean and variance, and repeat the obstacle-avoidance data collection with a clean train/test split of the 50 trajectories. If the LIBERO gap over BAKU falls to near zero or the intent prediction error rises to the level of the 1.94 mm baseline, the central claim would not survive.

Watch

Extended reading notes

Core claim

The central claim is that human intent estimation and multi-task policy learning can be fused into one architecture that autonomously infers which collaboration mode is active and adjusts robot actions accordingly. The model treats a collaboration episode as a conditional distribution of future human trajectories given past motion, force, vision, and language data, and uses the sampled latent intention to condition a Transformer decoder that outputs action chunks. The authors argue that adding force/torque data to intent estimation is what allows the robot to detect abrupt intention changes, such as a human pulling a guided arm away from an obstacle, and that the same latent variable supports smooth transitions between physical interaction and remote cooperation. The quantitative support is that the framework reaches state-of-the-art success rates on LIBERO-90 and LIBERO-10 while keeping the number of demonstrations small, and that its predicted trajectories are substantially closer to true human motion than existing short- and long-term predictors.

Load-bearing premise

The load-bearing premise is that the evaluation numbers are trustworthy: LIBERO success rates are reported as points without variance, and the real-world intent tests rely on a single obstacle-avoidance scenario whose data counts are internally inconsistent (50 collected trajectories but 80 train and 20 test). If those runs are few or the test trajectories are not truly held out, the headline gains over BAKU and the 0.26 mm error could be noise.

Editorial extensions

If this is right

  • A single policy can move between physical and remote collaboration modes without a hand-coded mode-switching rule, because the mode is inferred from the same latent intent that drives action generation.
  • Robot skills can be acquired from roughly 50 demonstrations per task, which is practical for real deployment where collecting physical interaction data is slow and costly.
  • Long-horizon tasks (LIBERO-10) can be learned with nearly the same success rate as short-horizon ones, suggesting that action chunking plus the controller hierarchy absorbs the added planning burden.
  • Force-aware intent prediction of about 0.26 mm average error could be used for early obstacle avoidance and safer physical handovers.

Reading between the lines

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

  • Going beyond the paper: the same CVAE intent module could be attached to any vision-language-action backbone, since it consumes standard trajectory and force sequences and outputs a conditioning vector.
  • A testable extension is to benchmark IDAGC on a suite with more than one obstacle configuration and repeated seeds; if the reported margins over BAKU shrink to noise on that suite, the advantage would likely come from the extra sensory channels rather than the architecture.
  • One untested consequence: a human deliberately changing direction mid-motion should produce an early jump in the latent variable's uncertainty, which could serve as a safety trigger for stopping or slowing the robot.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes IDAGC, a unified human-robot collaboration framework that combines a Conditional Variational Autoencoder (CVAE) based human intent encoder with a multimodal Transformer policy. The framework is designed to automatically switch between physical human-robot interaction (pHRI) and remote cooperation, using vision, language, force, and robot-state inputs. The paper reports state-of-the-art success rates on LIBERO-90 (91.1%) and LIBERO-10 (91.0%), a human intent prediction error of 0.26 mm compared with 1.94 mm for the best baseline, and qualitative real-world demonstrations of long-horizon tasks with mode switching.

Significance. If the reported results are reproducible, IDAGC would be a meaningful contribution: it integrates force-based intent estimation with language-conditioned multi-task policy learning in one architecture, extends action chunking and controller hierarchies to HRC, and addresses a real gap in mode-switching between physical and remote collaboration. The framework design is coherent and builds on established components (FiLM-conditioned ResNet, MiniLM text encoder, CVAE latent variable, Transformer decoder). However, the evidence supporting the central claims is thin: LIBERO results are point estimates without variance, the intent-estimation comparison contains an internally inconsistent data split, and the real-world mode-switching demonstrations lack baselines and quantitative switching metrics. The architectural novelty is promising, but the experimental validation does not yet meet the standard needed to substantiate the headline claims.

major comments (4)
  1. [IV-A.2] The data-collection description contains a direct numerical contradiction: the text states that "a total of 50 sets of data" were collected (25 per obstacle) and then says the trajectories "were divided into two datasets, among which 80 were for training, and 20 were for testing." A set of 50 trajectories cannot be partitioned into 80 and 20. If 80/20 are percentages, the text is misleading; if they are counts, the reported training/test split is impossible. This inconsistency directly affects the headline intent-prediction comparison in Table II (0.26 mm versus 1.94 mm), since the HIE is trained and evaluated on this split. The authors must clarify the exact number of training and test trajectories, and specify whether the split was performed at the trajectory level (never mixing frames from the same trajectory across train and test) to avoid leakage.
  2. [IV-A.1] Table I reports LIBERO success rates as single point values, with the text stating only that "All evaluations are conducted using 10 policies for each task." No standard deviations, number of evaluation rollouts per policy, or statistical significance tests are provided. The claimed improvements over BAKU are 1.2 percentage points on LIBERO-90 and 6.0 percentage points on LIBERO-10; without variance or rollout counts, these differences could easily be within noise. The authors should report mean and standard deviation over multiple seeds, the number of evaluation episodes per task, and ideally per-task confidence intervals, to support the state-of-the-art claim.
  3. [IV-C] The real-world long-horizon experiments are the only evidence for the central claim of autonomous collaboration-mode switching, but they are presented solely through qualitative descriptions, snapshots in Figs. 6-7, and result curves in Fig. 8. There is no baseline comparison (e.g., fixed-mode pHRI or remote-only policies), no quantitative metric for switching accuracy or switching latency, and no measure of how often the mode transition was correct. Without such evidence, the claim that IDAGC "autonomously infers collaboration modes" is not demonstrated. The authors should provide a quantitative evaluation of mode-switch decisions, such as precision/recall of detected transitions against a ground-truth labeling, or a comparative study against methods that do not switch modes.
  4. [IV (general)] The paper repeatedly refers to "supplementary material" for data collection, simulations, comparison experiments, and applications, but no supplementary document or repository is provided with the manuscript. Since the central experimental claims depend on details that are not in the main text (e.g., exact training hyperparameters, baseline configurations, number of real-world trials), the evaluation is not independently verifiable. The authors should either include the supplementary material or move the essential experimental details into the paper.
minor comments (5)
  1. [Abstract] "Experiments highlights" should be "Experiments highlight" for subject-verb agreement.
  2. [Fig. 4] The figure has two items labeled "(d)": one for the collected-trajectory samples and one for the visualization of predictions. Rename the second label (e.g., "(e)") and update the caption accordingly.
  3. [IV-B] The metric is referred to as "PHRI angle" in the main text but as "pHRC angle" in the caption of Fig. 5. Use one consistent term throughout.
  4. [III-A, Eq. (2)] Equation (2) is introduced as the "weighted negative evidence-based lower bound (ELBO)" but the displayed expression is L_ELBO + L_BC with positive weighting on the KL term and negative weight on the log-likelihood term. Calling it a weighted ELBO would be more accurate than "negative ELBO."
  5. [IV-A.2] The text says "In simulation, we train the proposed framework and validate the human intent estimation accuracy" and then immediately repeats a similar sentence; the duplication should be removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central HIE and policy results are validated against external benchmarks and baselines, and the few self-citations are not load-bearing.

full rationale

The paper's central derivation chain is not circular. The CVAE intent model in Eq. (1) and the ELBO-plus-behavior-cloning loss in Eq. (2) define a standard conditional generative objective; at test time the HIE samples from the past-only prior p(Z|X), so future trajectory predictions are not constructed from ground-truth future information. The multi-task policy claims are evaluated on the external LIBERO benchmark [35] against external baselines MT-ACT [8] and BAKU [12] (Table I, Fig. 3), and the pHRI intent comparison uses the external methods [13]-[15] (Table II). No fitted parameter is renamed as a prediction. The self-citations [2], [17], [23] appear only in introductory or related-work motivation (e.g., impedance/admittance control and limitations of motion-only long-term prediction) and do not supply the benchmark numbers, define the loss, or force the architectural choices. The reported '50 sets ... 80 were for training, and 20 were for testing' inconsistency and the lack of variances are evaluation-support or correctness concerns, not definitional circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claims rest on a small number of hand-set hyperparameters (loss weights, horizons, architecture choices) that are not reported, on domain assumptions about the sufficiency of small demonstration sets and sensor synchronization, and on one latent intent variable Z that has no independent evidence. None of these are fatal by themselves, but together they make the reported accuracies difficult to reproduce or interpret.

free parameters (4)
  • Loss weights alpha, beta, gamma in Eq. 2 = not reported
    Hand-chosen weights balancing ELBO and behavior cloning; no values or sensitivity analysis supplied.
  • CVAE latent dimension and architecture hyperparameters = not reported
    Number of latent dimensions, attention heads, and layers for the human intent encoder are omitted.
  • Training hyperparameters (learning rate, batch size, epochs, action chunk size) = not reported
    No training schedule is given; replication would require guessing.
  • Observation and prediction horizons Tobs, Tfut = not reported
    These define the CVAE input and output windows but their numeric values are not stated.
assumptions (5)
  • domain assumption Pretrained encoders (ResNet-18, MiniLM, Sentence Transformers) and borrowed architectures (BAKU, SERL) behave as expected when combined.
    The framework inherits the reliability of external components without validating their behavior in this new combination.
  • domain assumption Fifty demonstrations per LIBERO task and the collected pHRI trajectory sets are sufficient to train the multimodal policy and intent model.
    Few-shot generalization is assumed; no scaling, variance, or cross-validation evidence is given.
  • domain assumption Sensor streams (cameras, force/torque, robot state) are synchronized and calibrated.
    Multimodal fusion requires temporal alignment; the paper does not describe synchronization or calibration procedures.
  • domain assumption A Gaussian conditional prior over latent intention Z can represent the multimodal distribution of future trajectories.
    The CVAE formulation in Eq. 1 assumes this probabilistic model; no identifiability or goodness-of-fit analysis is provided.
  • domain assumption Success rate on LIBERO is a valid proxy for multi-task policy quality.
    Standard benchmark, but only the success metric is reported and no confidence intervals are given.
invented entities (1)
  • Latent intent variable Z
    purpose: Unobserved human intention used in Eq. 1 to condition future trajectory prediction and to drive collaboration mode transitions.
    Z is a fitted latent variable with no external measurable counterpart; the paper provides no falsifiable prediction tied to Z outside the trained model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IDAGC: Adaptive Generalized Human-Robot Collaboration via Human Intent Estimation and Multimodal Policy Learning." pith.science (2026). https://pith.science/paper/Z6E3HSPT

@misc{pith2026250704620,
  author       = {Pith},
  title        = {Pith review of: IDAGC: Adaptive Generalized Human-Robot Collaboration via Human Intent Estimation and Multimodal Policy Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z6E3HSPT}},
  note         = {Machine review of arXiv:2507.04620}
}
read the original abstract

In Human-Robot Collaboration (HRC), which encompasses physical interaction and remote cooperation, accurate estimation of human intentions and seamless switching of collaboration modes to adjust robot behavior remain paramount challenges. To address these issues, we propose an Intent-Driven Adaptive Generalized Collaboration (IDAGC) framework that leverages multimodal data and human intent estimation to facilitate adaptive policy learning across multi-tasks in diverse scenarios, thereby facilitating autonomous inference of collaboration modes and dynamic adjustment of robotic actions. This framework overcomes the limitations of existing HRC methods, which are typically restricted to a single collaboration mode and lack the capacity to identify and transition between diverse states. Central to our framework is a predictive model that captures the interdependencies among vision, language, force, and robot state data to accurately recognize human intentions with a Conditional Variational Autoencoder (CVAE) and automatically switch collaboration modes. By employing dedicated encoders for each modality and integrating extracted features through a Transformer decoder, the framework efficiently learns multi-task policies, while force data optimizes compliance control and intent estimation accuracy during physical interactions. Experiments highlights our framework's practical potential to advance the comprehensive development of HRC.

Figures

Figures reproduced from arXiv: 2507.04620 by the authors.

Figure 1
Figure 1. Overview of the IDAGC framework. the coupling characteristics of policies across multimodal information. Consequently, there is an urgent need for an adaptive framework that integrates remote cooperation with physical interaction through human-like intuition. In pHRI, accurate prediction of future trajectories based on human intention is essential for efficient robot assistance and safety assurance [5]. In complex e… view at source ↗
Figure 2
Figure 2. The structure of IDAGC. The green lines denote the train-only operations, while the blue lines denote the predict-only operations. The dashed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance of BAKU [12] and IDAGC on LIBERO-10 [35]. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Process and results of pHRI. (a) pHRI process. (b) Comparison [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Processes of placing tasks. (a) The robot grasps the white plug ( [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Process of writing task. The robot grasps the red marker ( [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Results of writing task. (a) Predictions of human intention and robot [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 34 canonical work pages

  1. [13]

    Model-free online neuroadaptive controller with intent estimation for physical human-robot interaction,

    S. Cremer, S. K. Das, I. B. Wijayasinghe, D. O. Popa, and F. L. Lewis, “Model-free online neuroadaptive controller with intent estimation for physical human-robot interaction,” IEEE Transactions on Robotics , vol. 36, no. 1, pp. 240–253, 2020

    work page 2020

  2. [15]

    Learning human motion intention for phri assistive control,

    P. Franceschi, F. Bertini, F. Braghin, L. Roveda, N. Pedrocchi, and M. Beschi, “Learning human motion intention for phri assistive control,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . Detroit, MI, USA: IEEE, 2023, pp. 7870–7877

    work page 2023

  3. [1]

    Impedance learning for human-guided robots in contact with unknown environments,

    X. Xing, E. Burdet, W. Si, C. Yang, and Y . Li, “Impedance learning for human-guided robots in contact with unknown environments,” IEEE Transactions on Robotics , vol. 39, no. 5, pp. 3705–3721, 2023

    work page 2023

  4. [2]

    Human observation-inspired uni- versal image acquisition paradigm integrating multi-objective motion planning and control for robotics,

    H. Liu, Y . Tong, and Z. Zhang, “Human observation-inspired uni- versal image acquisition paradigm integrating multi-objective motion planning and control for robotics,” IEEE/CAA Journal of Automatica Sinica, vol. 11, no. 12, pp. 2463–2475, 2024

    work page 2024

  5. [3]

    Adaptive- constrained impedance control for human-robot co-transportation,

    X. Yu, B. Li, W. He, Y . Feng, L. Cheng, and C. Silvestre, “Adaptive- constrained impedance control for human-robot co-transportation,” IEEE Transactions on Cybernetics, vol. 52, no. 12, pp. 13 237–13 249, 2022

    work page 2022

  6. [4]

    Towards proactive safe human-robot collaborations via data-efficient conditional behavior prediction,

    P. Ravi, W. Zhuoyuan, N. Yorie, and L. Changliu, “Towards proactive safe human-robot collaborations via data-efficient conditional behavior prediction,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . Yokohama, Japan: IEEE, 2024, pp. 12 956– 12 963

    work page 2024

  7. [5]

    Hierarchical human motion inten- tion prediction for increasing efficacy of human-robot collaboration,

    L. Meng, L. Yang, and E. Zheng, “Hierarchical human motion inten- tion prediction for increasing efficacy of human-robot collaboration,” IEEE Robotics and Automation Letters , vol. 9, no. 9, pp. 7637–7644, 2024

    work page 2024

  8. [6]

    Follow the force: Haptic com- munication enhances coordination in physical human-robot interaction when humans are followers,

    Y . Liu, R. Leib, and D. W. Franklin, “Follow the force: Haptic com- munication enhances coordination in physical human-robot interaction when humans are followers,” IEEE Robotics and Automation Letters , vol. 8, no. 10, pp. 6459–6466, 2023

    work page 2023

Show all 35 references
  1. [7]

    An online human dynamic arm strength perception method based on surface electromyography signals for human-robot collaboration,

    T. Zhang, H. Chu, and Y . Zou, “An online human dynamic arm strength perception method based on surface electromyography signals for human-robot collaboration,” IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–14, 2023

    2023

  2. [8]

    Roboagent: Generalization and efficiency in robot ma- nipulation via semantic augmentations and action chunking,

    H. Bharadhwaj, J. Vakil, M. Sharma, A. Gupta, S. Tulsiani, and V . Kumar, “Roboagent: Generalization and efficiency in robot ma- nipulation via semantic augmentations and action chunking,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . Yokohama, Japa...

    2024

  3. [9]

    Openvla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burch- fiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn, “Openvla: An open-source vision-language-action model,” in Proceeding...

    2024

  4. [10]

    Extract: Efficient policy learning by extracting transferable robot skills from offline data,

    J. Zhang, M. Heo, Z. Liu, E. Bıyık, J. J. Lim, Y . Liu, and R. Fakoor, “Extract: Efficient policy learning by extracting transferable robot skills from offline data,” in Proceedings of the 8th conference on robot learning. Munich, Germany: Robot Learning Foundation, 2024-11- 0...

    2024

  5. [11]

    Modem: accelerating visual model-based reinforcement learning with demonstrations,

    N. Hansen, Y . X. Lin, H. Su, X. L. Wang, V . Kumar, and A. Ra- jeswaran, “Modem: accelerating visual model-based reinforcement learning with demonstrations,” in The Eleventh International Confer- ence on Learning Representations , 2023

    2023

  6. [12]

    BAKU: An efficient transformer for multi-task policy learning,

    S. Haldar, Z. Peng, and L. Pinto, “BAKU: An efficient transformer for multi-task policy learning,” in Advances in Neural Information Processing Systems , vol. 37. Curran Associates, Inc., 2024, pp. 141 208–141 239

    2024

  7. [14]

    A human-robot collaboration controller utiliz- ing confidence for disagreement adjustment,

    M. Ma and L. Cheng, “A human-robot collaboration controller utiliz- ing confidence for disagreement adjustment,” IEEE Transactions on Robotics, vol. 40, pp. 2081–2097, 2024

    2024

  8. [16]

    Constraint-aware intent estimation for dynamic human-robot object co-manipulation,

    Y . Shao, T. Li, S. Keyvanian, P. Chadhuari, V . Kumar, and N. Figueroa, “Constraint-aware intent estimation for dynamic human-robot object co-manipulation,” in Proceedings of Robotics: Science and Systems , Delft, Netherlands, July 2024

    2024

  9. [17]

    Human-inspired adaptive optimal con- trol framework for robot-environment interaction,

    H. Liu, Y . Tong, and Z. Zhang, “Human-inspired adaptive optimal con- trol framework for robot-environment interaction,” IEEE Transactions on Systems, Man, and Cybernetics: Systems , pp. 1–14, 2025

    2025

  10. [18]

    Role dynamic assignment of human-robot collaboration based on target prediction and fuzzy inference,

    C. Wang and J. Zhao, “Role dynamic assignment of human-robot collaboration based on target prediction and fuzzy inference,” IEEE Transactions on Industrial Informatics , vol. 20, no. 1, pp. 471–481, 2024

    2024

  11. [19]

    Shared navigational control and user intent detection in an intelligent walker,

    C. Huang, G. S. Wasson, M. Alwan, P. Sheth, and A. Ledoux, “Shared navigational control and user intent detection in an intelligent walker,” in AAAI Fall Symposium: Caring Machines , 2005

    2005

  12. [20]

    A characterization of the effect of limb position on emg features to guide the development of effective prosthetic control schemes,

    A. Radmand, E. Scheme, and K. Englehart, “A characterization of the effect of limb position on emg features to guide the development of effective prosthetic control schemes,” in 2014 36th Annual Interna- tional Conference of the IEEE Engineering in Medicine and Biology Society...

    2014

  13. [21]

    Hybrid recurrent neu- ral network architecture-based intention recognition for human-robot collaboration,

    X. Gao, L. Yan, G. Wang, and C. Gerada, “Hybrid recurrent neu- ral network architecture-based intention recognition for human-robot collaboration,” IEEE Transactions on Cybernetics , vol. 53, no. 3, pp. 1578–1586, 2023

    2023

  14. [22]

    Robot trajectron: Trajec- tory prediction-based shared control for robot manipulation,

    P. Song, P. Li, E. Aertbelien, and R. Detry, “Robot trajectron: Trajec- tory prediction-based shared control for robot manipulation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . Yokohama, Japan: IEEE, 2024

    2024

  15. [23]

    DTRT: Enhancing human intent estimation and role allocation for physical human-robot collaboration,

    H. Liu, Y . Tong, and Z. Zhang, “DTRT: Enhancing human intent estimation and role allocation for physical human-robot collaboration,” in 2025 IEEE International Conference on Robotics and Automation (ICRA). Atlanta, USA: IEEE, 2025

    2025

  16. [24]

    Learn to grasp via intention discovery and its application to challenging clutter,

    C. Zhao, C. L. Jiang, J. H. Cai, H. Y . Yu, M. Y . Wang, and Q. F. Chen, “Learn to grasp via intention discovery and its application to challenging clutter,” IEEE Rob. Autom. Lett. , vol. 8, no. 2, pp. 488– 495, 2023

    2023

  17. [25]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” in Proceedings of Robotics: Science and Systems , Daegu, Republic of Korea, 2023

    2023

  18. [26]

    Perceiver-actor: a multi- task transformer for robotic manipulation,

    M. Shridhar, L. Manuelli, and D. Fox, “Perceiver-actor: a multi- task transformer for robotic manipulation,” in Proceedings of The 6th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, vol. 205. Proceedings of The 6th Conference on Robot Learning, 20...

    2023

  19. [27]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” in Proceedings of the 38th International Conference on Machine...

    2021

  20. [28]

    Sigmoid loss for language image pre-training,

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer, “Sigmoid loss for language image pre-training,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , 2023, pp. 11 941–11 952

    2023

  21. [29]

    Adaptive compliance policy: Learning approximate compliance for diffusion guided control,

    Y . Hou, Z. Liu, C. Chi, E. Cousineau, N. Kuppuswamy, S. Feng, B. Burchfiel, and S. Song, “Adaptive compliance policy: Learning approximate compliance for diffusion guided control,” in 2025 IEEE International Conference on Robotics and Automation (ICRA) . At- lanta, USA: IEEE, 2025

    2025

  22. [30]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2016, pp. 770–778

    2016

  23. [31]

    Minilm: deep self-attention distillation for task-agnostic compression of pre- trained transformers,

    W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, and M. Zhou, “Minilm: deep self-attention distillation for task-agnostic compression of pre- trained transformers,” in Advances in neural information processing systems, vol. 33. Curran Associates, Inc., 2020, pp. 5776–5788

    2020

  24. [32]

    Sentence-BERT: Sentence embed- dings using Siamese BERT-networks,

    N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embed- dings using Siamese BERT-networks,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). H...

    2019

  25. [33]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. C. Courville, “Film: Visual reasoning with a general conditioning layer,” inProceed- ings of the Thirty-Second AAAI Conference on Artificial Intelligence . AAAI Press, 2018, pp. 3942–3951

    2018

  26. [34]

    Serl: A software suite for sample- efficient robotic reinforcement learning,

    J. Luo, Z. Hu, C. Xu, Y . L. Tan, J. Berg, A. Sharma, S. Schaal, C. Finn, A. Gupta, and S. Levine, “Serl: A software suite for sample- efficient robotic reinforcement learning,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . Yokohama, Japan: IEEE, 20...

    2024

  27. [35]

    Libero: benchmarking knowledge transfer for lifelong robot learning,

    B. Liu, Y . F. Zhu, C. K. Gao, Y . H. Feng, Q. Liu, Y . K. Zhu, and P. Stone, “Libero: benchmarking knowledge transfer for lifelong robot learning,” in Advances in Neural Information Processing Systems , vol. 36. Advances in Neural Information Processing Systems, 2023, pp. 44 ...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.