Pith. sign in

REVIEW 4 major objections 5 minor 56 references

Probabilistic Human Intent Prediction for Mobile Manipulation: An Evaluation with Human-Inspired Constraints

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 The paper claims a robot can infer a teleoperator's intent as two coupled probabilistic beliefs—where the base is heading and which object the arm will grasp—without any predefined goal list, and that this coupling plus grasp-feasibility…

desk verdict The unified dual-phase intent framework is a real contribution, but the headline 'three times earlier' result is a designed consequence of the feasibility cascade, not a clean measure of intent recognition. read the letter →

arxiv 2507.10131 v1 pith:SSESWZTT submitted 2025-07-14 cs.RO cs.CVcs.HC

Cesar Alan Contreras , Manolis Chiou , Alireza Rastegarpanah , Michal Szulik , Rustam Stolkin This is my paper
classification cs.ROcs.CVcs.HC
keywords humanintentpredictionmobilemanipulationprobabilisticbeliefteleoperationgraspfeasibilitysaliencyvariableautonomyhuman-robotinteraction
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

GUIDER claims that intent in teleoperated mobile manipulation is best tracked as two coupled belief layers: one over navigation areas and one over graspable objects, instead of two independent predictors. In the navigation phase, a Synergy Map fuses the operator's velocity commands with the occupancy map so that a hot spot emerges at the likely interaction area with no predefined waypoints. After an autonomous 3D scan, the manipulation phase scores candidate objects through saliency, instance segmentation, three geometric grasp-feasibility tests, and a kinematics-aware update that grows or shrinks each object's probability as the gripper moves. Across 25 simulated trials, the paper reports higher median prediction stability than both baselines (93–100% versus 60–100% in navigation; 94–100% versus 69–100% in manipulation), and in the geometry-constrained task it identifies the correct graspable object at a median 23.6 seconds before contact compared with 7.8 seconds for Trajectron. A sympathetic reading is that a single coupled belief structure with physical-feasibility checks is what makes intent predictions both stable and early.

What carries the argument

The carrying mechanism is the coupled two-layer belief: in navigation, $B_{\text{combined}}(x,y,t) = \max\{B_{\text{base}}, B_{\text{motion}}, B_{\text{synergy}}\}$ forms a Synergy Map whose peak indicates the intended interaction area; in manipulation, a pixel-level cascade multiplies saliency by a center bias, geometric-feasibility masks, and depth weighting, pools the result into object scores, and then evolves each object's probability with a kinematics-aware growth-decay model. The named objects are the Synergy Map and the End-Effector Probability Evolution rule, and the work they do is to turn operator commands and scene geometry into a smooth, prioritized belief that avoids sudden goal flips.

What would settle it

Feed Robot Trajectron an object list that is independent of GUIDER's perception module, for example ground-truth object centroids or objects from an external detector, and rerun the five manipulation tasks; if the median remaining-time gap (23.6 s versus 7.8 s) collapses, then the reported timing advantage is an artifact of shared perception rather than a property of the coupled belief and geometry checks.

Watch

Extended reading notes

Core claim

The core claim is that human operator intent during mobile manipulation can be continuously estimated as a dual-phase probabilistic belief: an area-level map belief for where the base is heading and an object-level belief for what the end effector will grasp, with the two layers sharing information downstream. During navigation, the Synergy Map combines a decaying base layer, a motion-evidence layer built from multi-horizon velocity projections, and a flood-fill synergy layer, and the map's peak is taken as the predicted interaction area without a predefined goal set. On arrival, a roughly 35-second autonomous multi-view scan builds a fused local 3D cloud. The manipulation phase then fuses U2-Net saliency with FastSAM instance masks prompted by HDBSCAN centroid clusters, applies three grasp-feasibility tests (bounding-box, morphological erosion, and an advanced contour-pair check), and runs a forward-Euler update where object probabilities grow with proximity and approach speed and decay as the gripper moves away. The paper argues this design mirrors human spatial-to-focal reasoning and that its geometry checks are what let the system detect grasp-infeasible targets early.

Load-bearing premise

The load-bearing premise is that both baseline methods were given fairly comparable inputs, with the navigation baseline receiving the same four target regions and the manipulation baseline receiving the object list produced by the proposed system's own perception module; if those inputs are not equivalent, the reported differences could reflect implementation asymmetry rather than a genuine advantage.

Editorial extensions

If this is right

  • If GUIDER is right, a single coupled belief over base and arm can hold correct intent through both navigation redirections and manipulation redirections, reducing erratic predictions where separate predictors flip.
  • The geometric-feasibility cascade means the robot can rule out grasp-infeasible objects before the operator finishes reaching, which is why the paper finds the largest timing gain in the geometry-constrained task.
  • Goal-free inference would let intent emerge during operation, removing the need to pre-enumerate waypoints or candidate objects, which is useful in unstructured environments.
  • The framework's high-stability belief can serve as the inference backbone for variable-autonomy controllers, which could take over base motion or grasp execution once intent confidence crosses a threshold.
  • Only the geometry-aware ranking contributed the reported quantitative gains; the other design features enable goal-free area and object identification and smooth redirection handling.

Reading between the lines

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

  • If the timing advantage is robust, the grasp-feasibility cascade could be detached and added to any trajectory-based manipulation predictor, which would test whether the gain comes from affordance gating rather than from GUIDER's perception module.
  • The no-predefined-goals property is tested only within four fixed regions and a small object set; a harder test would place the robot in an open room with never-seen objects and measure whether the synergy map and saliency cascade still converge.
  • A shared-control controller could use the predicted object's probability as a confidence signal to decide when to offer assistance, for example triggering autonomous grasp execution only when the probability stays high for a sustained interval.
  • The hyperparameters were manually tuned on pilot data and kept fixed; an adaptive version that updates decay and growth rates online might transfer better to real sensor noise, but that is not established by the current study.
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 paper presents GUIDER, a dual-phase probabilistic framework for estimating a teleoperator's intent during mobile manipulation. In the navigation phase, a Synergy Map combines controller velocity with an occupancy grid to rank interaction areas; after an autonomous 3D scan, the manipulation phase fuses U2-Net and FastSAM saliency with three geometric grasp-feasibility tests and an end-effector kinematics-aware update rule to rank candidate objects. The authors evaluate the framework post-hoc on 25 simulated trials (five participants by five task variants) in Isaac Sim, comparing against BOIR for navigation and Robot Trajectron for manipulation. They report higher prediction stability for GUIDER in both phases and, in the geometry-constrained T5 task, a median remaining-time-to-confident-prediction of 23.6 s versus 7.8 s for Trajectron, which the abstract summarizes as recognizing object intent three times earlier.

Significance. If the reported effects were cleanly measured, GUIDER would be a useful system-level contribution: it couples navigation and manipulation beliefs in one structure, avoids predefined goal lists, and incorporates grasp feasibility directly into intent scoring. The paper also has strengths in transparency: the method is specified in detail with equations and hyperparameters, the hypotheses are stated before reporting, and the authors openly acknowledge the simulation-only, five-participant evaluation and manual hyperparameter tuning. However, the headline quantitative claim is confounded by construction: the T5 RTCP advantage is driven by the built-in grasp-feasibility filter rather than by evidence that the system recognized the operator's current intent, and the Robot Trajectron comparison is incompletely specified. The broader dual-phase claims are further weakened by the paper's own statement that only the geometric-feasibility feature influenced the quantitative scores. The framework may still be a valuable proof of concept, but the current evaluation does not support the abstract's strongest conclusion.

major comments (4)
  1. [Sections 3.3.6-3.3.7, Table 2, Section 5] The headline 'three times earlier' result in T5 is built into the method rather than measuring intent recognition. The geometric feasibility cascade in Section 3.3.6 and the mask multiplication with lambda_in=1.2 and lambda_out=0.4 in Section 3.3.7 cause GUIDER to down-weight the ungraspable coffee can and rank the feasible fruit before any operator motion, and Section 5 confirms that the food object was ranked first in all five T5 trials. Because RTCP is measured against the final target (the fruit) rather than the currently instructed target (the coffee can during step iii), the earlier 'confident prediction' reflects affordance filtering rather than recognition of the operator's present intent. The conclusion in Section 7 that the three-times advantage is 'a direct result of its built-in grasp affordance analysis' is consistent with this reading, and it should be stated as such in the abstract rather than presented as earlier intent recognition.
  2. [Section 4.2] The comparison with Robot Trajectron is not interpretable as an intent-recognition comparison. Trajectron is given only the object list returned by GUIDER's perception module, while GUIDER additionally receives the feasibility masks and the kinematics-based end-effector update; the conversion of Trajectron's trajectory predictions into object-level intent probabilities is not documented. Without a precise description of that conversion, and without matching the information available to each method, the reported RTCP and stability differences could reflect an implementation asymmetry. Please specify the conversion algorithm, or preferably run Trajectron with the same perceptual inputs and object-scoring procedure, and report the effect of removing the feasibility masks from GUIDER as an ablation.
  3. [Section 6] The statement 'Only feature (c) influenced quantitative scores in our offline study' directly conflicts with the abstract's claim that 'These results validate our dual-phase framework.' The experimental design does not isolate the contribution of the coupled navigation-manipulation belief, goal-free inference, or the end-effector kinematics update; the reported quantitative improvements are attributed to the geometric-feasibility channel. The authors should either provide an ablation that isolates the dual-phase and goal-free components or substantially reframe the quantitative claims so that they are limited to the geometry-aware ranking feature.
  4. [Sections 3.3.7 and 3.4] The term 'probability' is used for scores that are not normalized over the object set. Each connected component receives gi = min P*(u,v) and is then linearly rescaled to [0.5, 0.7] independently, and Equations (7)-(8) evolve each object's value with clipping but no renormalization, so the values do not form a probability distribution over candidate objects. Because the framework is described as probabilistic and the RTCP and stability definitions rely on 'highest predicted probability', the authors should either introduce proper normalization or explicitly describe these values as heuristic belief scores rather than probabilities.
minor comments (5)
  1. [Table 2 and Section 4.1] Task T4 is described in Section 4.1 as requiring navigation to R1 (the tool pallet), but Table 2 labels the navigation-phase T4 row 'Tool grasp (R2)'; please correct the inconsistency.
  2. [Table 2, Pooled Stability row] The pooled manipulation stability row reports Trajectron as 100.0 +/- 0.0 even though the T3 row shows 68.6% and the T4 row shows 87.2%; please clarify how the pooled median and MAD are computed across trials and tasks.
  3. [Section 4.2] The text calls p1 the 'pre-registered' one-tailed test, but no pre-registration is mentioned anywhere else in the manuscript; please either provide a registration identifier or replace 'pre-registered' with 'one-tailed'.
  4. [Section 3.4] Equation (8) uses both pmax,I and pmax,i, and the top-K set TK(t) is used in the definition of the bias beta_i before K is fully specified; please make the notation and the order of definitions consistent.
  5. [Section 4.1 and Table 2] T3 is called 'Manipulator Redirection' but it still includes a navigation phase to R3; please clarify in the task description and in Table 2 that the navigation-phase row for T3 refers to the initial drive to R3.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline T5 'three-times-earlier' result reduces by construction: feasibility masks are an input that pre-boosts the feasible object, and RTCP is scored against the final target while the operator is still attempting the infeasible one.

  1. fitted input called prediction [Section 3.3.7 (mask multiplication), Section 4.1 (T5 task definition), Section 5 (T5 result)]
    "P(k+1)(u, v) = λin P(k)(u, v), M(u, v) = 1, λout P(k)(u, v), M(u, v) = 0, with λin = 1.2 and λout = 0.4. Pixels consistent with a feasible-geometry mask are boosted; pixels that don’t appear on a mask are de-emphasised but not nulled. ... In the geometrically constrained scenario (T5), GUIDER ranked the food object first in all five trials. ... T5 ... (iii) Try to grasp the large coffee can (end-effector too small). (iv) Receive a final cue to grasp the adjacent fruit."

    The T5 ground-truth target is the fruit, reached only after the operator first tries the infeasible coffee can. The feasibility masks are computed before teleoperation resumes, and the mask multiplication (λin=1.2 / λout=0.4) boosts the fruit while down-weighting the can. Thus GUIDER's top-ranked 'prediction' at the moment the manipulation phase starts is the output of its own geometric filter, not an inference from operator motion or current intent. RTCP measures from that pre-assigned top rank, so the reported 23.6 s vs 7.8 s advantage is the feasibility cascade by construction.

full rationale

The overall GUIDER framework is not circular in its architecture: the Synergy Map, saliency fusion, and kinematics-based probability evolution are all specified from external perceptual and motion inputs, and the stability results are empirical comparisons against Trajectron and BOIR. No load-bearing uniqueness theorem or self-citation chain was found; BOIR is a self-cited prior method but is used as an ordinary baseline, not as justification for GUIDER's validity. However, the paper's headline manipulation claim—'recognized the object intent three times earlier than Trajectron'—is partially circular. In T5, the perceptual cascade's feasibility masks are an input that, by construction, gives the fruit the top initial score; RTCP then credits that pre-assigned score as an early correct prediction against a ground-truth label that is the final instructed target, even during the phase when the operator is attempting the infeasible coffee can. The paper's own wording ('due to our integrated geometric feasibility constraints', 'a direct result of its built-in grasp affordance analysis') confirms that the RTCP gap is the feasibility filter re-stated as intent prediction. Separately, the baseline comparison is hard to interpret because Robot Trajectron receives the object list but not the feasibility masks (Section 4.2), which is an information asymmetry rather than a circularity itself. Because the central 'three-times-earlier' claim reduces by construction while the rest of the system has independent content, a score of 6 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

GUIDER introduces no new physical entities, forces, or dimensions. Its burden comes instead from a large set of hand-tuned constants (Table 1 and the manipulation-phase equations) and from domain assumptions about the simulator, the ground-truth labels, the probabilistic interpretation of unnormalized scores, and the Synergy Map's meaning.

free parameters (5)
  • Navigation hyperparameters (Table 1) = Delta_cell=0.05 m; gamma_base_free=0.55; gamma_base_obj=0.50; r_infl=0.75 m; Delta_t_pred=0.10 s; d_upd=0.30 m…
    All 16 values in Table 1 are stated to be manually tuned and fixed; they directly shape the Synergy Map and therefore the navigation intent predictions.
  • Motion blend weights alpha_i = alpha = {0.60, 0.60, 0.85}
    Equation (2) blends three extrapolation horizons with weights that favor the farthest projection during steady driving; these hand-chosen weights affect where the motion evidence layer peaks.
  • Perceptual fusion constants = U2-Net threshold tau=0.9; FastSAM confidence threshold tau_conf=0.4; saliency weights 0.6/0.9; center sigma_c=240 px…
    These constants control how visual saliency, center bias, depth weighting, and geometric masks are combined into object scores; they are empirically chosen and not learned.
  • Geometric feasibility constants = gripper radius r=0.0425 m; finger thickness w_f=8.25 mm; d_skip=1 px; depth distance d_min=0.30 m; RANSAC distance…
    These hand-set geometry parameters determine which objects pass the bounding-box, morphological, and advanced grasp-feasibility tests, directly affecting the T5 RTCP result.
  • End-effector probability evolution constants = alpha_g=0.08; alpha_d=0.80; gamma_v=0.10; gamma_a=0.05; delta=0.10 m; kappa=10; p_cap=0.30; beta_i=0.002/0.005; K=2…
    Equations (5)-(8) use these experimentally tuned constants to grow and decay object probabilities; they determine the hysteresis and stability that the paper highlights as a key improvement.
assumptions (4)
  • ad hoc to paper Object scores are treated as probabilities without normalization across objects.
    In Section 3.4, Equations (7)-(8) update each object's score independently with clipping to [0, 0.99] and no normalization, so the scores do not form a proper probability distribution over the object set.
  • domain assumption The Synergy Map peak indicates the operator's next interaction area.
    Section 3.1 assumes that the pointwise maximum of base, motion, and synergy layers, B_combined, identifies where the operator will interact next; this is a modeling assumption, not a proven property.
  • domain assumption Manually annotated ground-truth labels from task instructions are correct.
    Section 4.2 scores predictions against manually annotated ground-truth labels generated from known task instructions; any labeling error would bias the RTCP and stability metrics.
  • domain assumption Isaac Sim teleoperation behavior is representative of real teleoperation for intent inference.
    The evaluation is entirely in the Isaac Sim simulator with five participants and a specific controller interface; the paper cites sim-to-real transfer works but provides no real-world validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Human Intent Prediction for Mobile Manipulation: An Evaluation with Human-Inspired Constraints." pith.science (2026). https://pith.science/paper/SSESWZTT

@misc{pith2026250710131,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Human Intent Prediction for Mobile Manipulation: An Evaluation with Human-Inspired Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SSESWZTT}},
  note         = {Machine review of arXiv:2507.10131}
}
read the original abstract

Accurate inference of human intent enables human-robot collaboration without constraining human control or causing conflicts between humans and robots. We present GUIDER (Global User Intent Dual-phase Estimation for Robots), a probabilistic framework that enables a robot to estimate the intent of human operators. GUIDER maintains two coupled belief layers, one tracking navigation goals and the other manipulation goals. In the Navigation phase, a Synergy Map blends controller velocity with an occupancy grid to rank interaction areas. Upon arrival at a goal, an autonomous multi-view scan builds a local 3D cloud. The Manipulation phase combines U2Net saliency, FastSAM instance saliency, and three geometric grasp-feasibility tests, with an end-effector kinematics-aware update rule that evolves object probabilities in real-time. GUIDER can recognize areas and objects of intent without predefined goals. We evaluated GUIDER on 25 trials (five participants x five task variants) in Isaac Sim, and compared it with two baselines, one for navigation and one for manipulation. Across the 25 trials, GUIDER achieved a median stability of 93-100% during navigation, compared with 60-100% for the BOIR baseline, with an improvement of 39.5% in a redirection scenario (T5). During manipulation, stability reached 94-100% (versus 69-100% for Trajectron), with a 31.4% difference in a redirection task (T3). In geometry-constrained trials (manipulation), GUIDER recognized the object intent three times earlier than Trajectron (median remaining time to confident prediction 23.6 s vs 7.8 s). These results validate our dual-phase framework and show improvements in intent inference in both phases of mobile manipulation tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 49 canonical work pages

  1. [1]

    Fron- tiers in Robotics and AI 12, 1540476 (2025)

    Contreras, C.A., Rastegarpanah, A., Chiou, M., Stolkin, R.: A mini-review on mobile manipulators with variable autonomy. Fron- tiers in Robotics and AI 12, 1540476 (2025)

    work page 2025

  2. [2]

    arXiv preprint arXiv:2409.00215 (2024)

    Shao, Y.S., Li, T., Keyvanian, S., Chaudhari, P., Kumar, V., Figueroa, N.: Constraint- aware intent estimation for dynamic human- robot object co-manipulation. arXiv preprint arXiv:2409.00215 (2024)

    arXiv 2024

  3. [3]

    Progress in Biomedical Engineering 4(3), 032004 (2022)

    Li, Y., Sena, A., Wang, Z., Xing, X., Babiˇ c, J., Asseldonk, E., Burdet, E.: A review on interaction control for contact robots through intent detection. Progress in Biomedical Engineering 4(3), 032004 (2022)

    work page 2022

  4. [4]

    In: 2016 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS), pp

    Chiou, M., Stolkin, R., Bieksaite, G., Hawes, N., Shapiro, K.L., Harrison, T.S.: Experimen- tal analysis of a variable autonomy frame- work for controlling a remotely operating mobile robot. In: 2016 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS), pp. 3581–3588 (2016). IEEE

    work page 2016

  5. [5]

    IEEE Transactions on Neural Systems and Rehabilitation Engineering 28(6), 1497–1506 (2020)

    Gopinath, D.E., Argall, B.D.: Active intent disambiguation for shared control robots. IEEE Transactions on Neural Systems and Rehabilitation Engineering 28(6), 1497–1506 (2020)

    work page 2020

  6. [6]

    In: 2021 30th IEEE International Conference on Robot & Human Interactive Communication (RO-MAN), pp

    Oh, Y., Sch¨ afer, T., R¨ uther, B., Toussaint, M., Mainprice, J.: A system for traded con- trol teleoperation of manipulation tasks using intent prediction from hand gestures. In: 2021 30th IEEE International Conference on Robot & Human Interactive Communication (RO-MAN), pp. 503–508 (2021). IEEE

    work page 2021

  7. [7]

    IEEE/ASME Transactions on Mecha- tronics 27(6), 4462–4474 (2022)

    Luo, S., Cheng, M., Ding, R., Wang, F., Xu, B., Chen, B.: Human–robot shared control based on locally weighted intent prediction for a teleoperated hydraulic manipulator sys- tem. IEEE/ASME Transactions on Mecha- tronics 27(6), 4462–4474 (2022)

    work page 2022

  8. [8]

    In: 2021 IEEE International Confer- ence on Robotics and Automation (ICRA), pp

    Wu, J., Sun, X., Zeng, A., Song, S., Rusinkiewicz, S., Funkhouser, T.: Spatial intention maps for multi-agent mobile manip- ulation. In: 2021 IEEE International Confer- ence on Robotics and Automation (ICRA), pp. 8749–8756 (2021). IEEE

    work page 2021

Show all 56 references
  1. [9]

    In: 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp

    Rothfuß, S., Chiou, M., Inga, J., Hohmann, S., Stolkin, R.: A negotiation-theoretic frame- work for control authority transfer in mixed-initiative robotic systems. In: 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp. 921–928 (2022). https://doi.or...

    2022

  2. [10]

    In: Companion of the 2021 ACM/IEEE International Conference 19 on Human-robot Interaction, pp

    Ramesh, A., Chiou, M., Stolkin, R.: Robot vitals and robot health: An intuitive approach to quantifying and communicating predicted robot performance degradation in human-robot teams. In: Companion of the 2021 ACM/IEEE International Conference 19 on Human-robot Interaction, pp...

    2021

  3. [11]

    In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

    Chiou, M., Epsimos, G.-T., Nikolaou, G., Pappas, P., Petousakis, G., M¨ uhl, S., Stolkin, R.: Robot-assisted nuclear disaster response: Report and insights from a field exercise. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4545–4552...

    2022

  4. [12]

    In: 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp

    Panagopoulos, D., Petousakis, G., Stolkin, R., Nikolaou, G., Chiou, M.: A bayesian- based approach to human operator intent recognition in remote mobile robot naviga- tion. In: 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp. 125–131 (2021). IEEE

    2021

  5. [13]

    The International Journal of Robotics Research 37(7), 717–742 (2018)

    Javdani, S., Admoni, H., Pellegrinelli, S., Srinivasa, S.S., Bagnell, J.A.: Shared auton- omy via hindsight optimization for teleopera- tion and teaming. The International Journal of Robotics Research 37(7), 717–742 (2018)

    2018

  6. [14]

    In: 2011 IEEE International Conference on Robotics and Automation, pp

    Ding, H., Reißig, G., Wijaya, K., Bortot, D., Bengler, K., Stursberg, O.: Human arm motion modeling and long-term prediction for safe and efficient human-robot-interaction. In: 2011 IEEE International Conference on Robotics and Automation, pp. 5875–5880 (2011). IEEE

    2011

  7. [15]

    Journal of Intelligent and Robotic Systems 45, 31–52 (2006)

    Tahboub, K.A.: Intelligent human-machine interaction based on dynamic bayesian net- works probabilistic intention recognition. Journal of Intelligent and Robotic Systems 45, 31–52 (2006)

    2006

  8. [16]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Song, P., Li, P., Aertbeli¨ en, E., Detry, R.: Robot trajectron: Trajectory prediction- based shared control for robot manipulation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 5585– 5591 (2024). IEEE

    2024

  9. [17]

    In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pp

    Salzmann, T., Ivanovic, B., Chakravarty, P., Pavone, M.: Trajectron++: Dynamically- feasible trajectory forecasting with hetero- geneous data. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pp. 683–700 (202...

    2020

  10. [18]

    Philosophical Transactions of the Royal Society B: Biological Sciences 362(1485), 1585–1599 (2007)

    Cisek, P.: Cortical mechanisms of action selection: the affordance competition hypoth- esis. Philosophical Transactions of the Royal Society B: Biological Sciences 362(1485), 1585–1599 (2007)

    2007

  11. [19]

    In: Proceedings of the 1st International Workshop on Advances in Service Robotics (ASER’03), pp

    Vanhooydonck, D., Demeester, E., Nuttin, M., Van Brussel, H.: Shared control for intel- ligent wheelchairs: an implicit estimation of the user intention. In: Proceedings of the 1st International Workshop on Advances in Service Robotics (ASER’03), pp. 176–182 (2003)

    2003

  12. [20]

    EPJ Nuclear Sciences & Technologies 8, 20 (2022)

    Santos, P., Recknagel, J., Knuth, M., Stein- bacher, K., Ritz, M., Wassmann, B., Fellner, D.: Robbe–robot-aided processing of assem- blies during the dismantling of nuclear power plants. EPJ Nuclear Sciences & Technologies 8, 20 (2022)

    2022

  13. [21]

    In: 2016 International Conference on Robotics and Automation for Humanitarian Applications (RAHA), pp

    Marturi, N., Rastegarpanah, A., Takahashi, C., Adjigble, M., Stolkin, R., Zurek, S., Kopicki, M., Talha, M., Kuo, J.A., Bekiroglu, Y.: Towards advanced robotic manipulation for nuclear decommissioning: A pilot study on tele-operation and autonomy. In: 2016 International Confer...

    2016

  14. [22]

    ATW International Journal for Nuclear Power 69, 38–44 (2024)

    Wong, C., Chapman, H., Lawton, S., Szu- lik, M., Tuck, O., Richardson, J.-P.: A pro- posed pathway to improve reliability claims for robotic systems in nuclear decommission- ing safety cases. ATW International Journal for Nuclear Power 69, 38–44 (2024)

    2024

  15. [23]

    IEEE Trans- actions on Industrial Electronics 63(9), 5964–5975 (2016)

    Rahman, S.M., Ikeura, R.: Weight- prediction-based predictive optimal position and force controls of a power assist robotic system for object manipulation. IEEE Trans- actions on Industrial Electronics 63(9), 5964–5975 (2016)

    2016

  16. [24]

    In: 2015 20 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pp

    Best, G., Fitch, R.: Bayesian intention inference for trajectory prediction with an unknown goal destination. In: 2015 20 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pp. 5817– 5823 (2015). IEEE

    2015

  17. [25]

    In: 2021 IEEE International Conference on Robotics and Automation (ICRA), pp

    Yoon, H., Sankaranarayanan, S.: Predictive runtime monitoring for mobile robots using logic-based bayesian intent inference. In: 2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 8565–8571 (2021). IEEE

    2021

  18. [26]

    Robotics and Autonomous Systems 56(8), 692–705 (2008)

    Aarno, D., Kragic, D.: Motion intention recognition in robot assisted applications. Robotics and Autonomous Systems 56(8), 692–705 (2008)

    2008

  19. [27]

    In: Proceedings of the 3rd ACM/IEEE International Con- ference on Human Robot Interaction, pp

    Kelley, R., Tavakkoli, A., King, C., Nicolescu, M., Nicolescu, M., Bebis, G.: Understanding human intentions via hidden markov models in autonomous mobile robots. In: Proceedings of the 3rd ACM/IEEE International Con- ference on Human Robot Interaction, pp. 367–374 (2008)

    2008

  20. [28]

    In: 2013 IEEE/RSJ International Conference on Intel- ligent Robots and Systems, pp

    Mainprice, J., Berenson, D.: Human-robot collaborative manipulation planning using early prediction of human motion. In: 2013 IEEE/RSJ International Conference on Intel- ligent Robots and Systems, pp. 299–306 (2013). IEEE

    2013

  21. [29]

    In: European Conference on Computer Vision, pp

    Yue, J., Manocha, D., Wang, H.: Human tra- jectory prediction via neural social physics. In: European Conference on Computer Vision, pp. 376–394 (2022). Springer

    2022

  22. [30]

    Neural computation 25(2), 328–373 (2013)

    Ijspeert, A.J., Nakanishi, J., Hoffmann, H., Pastor, P., Schaal, S.: Dynamical move- ment primitives: learning attractor models for motor behaviors. Neural computation 25(2), 328–373 (2013)

    2013

  23. [31]

    IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 42(4), 1039–1052 (2012)

    Vakanski, A., Mantegh, I., Irish, A., Janabi- Sharifi, F.: Trajectory learning for robot programming by demonstration using hid- den markov model and dynamic time warp- ing. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 42(4), 1039–1052 (2012)

    2012

  24. [32]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Ivanovic, B., Pavone, M.: The trajectron: Probabilistic multi-agent trajectory model- ing with dynamic spatiotemporal graphs. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 2375– 2384 (2019)

    2019

  25. [33]

    : Motiondif- fuser: Controllable multi-agent motion pre- diction using diffusion

    Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al. : Motiondif- fuser: Controllable multi-agent motion pre- diction using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9644– 9653 (2023)

    2023

  26. [34]

    In: 2021 IEEE International Confer- ence on Robotics and Automation (ICRA), pp

    Mandikal, P., Grauman, K.: Learning dexter- ous grasping with object-centric visual affor- dances. In: 2021 IEEE International Confer- ence on Robotics and Automation (ICRA), pp. 6169–6176 (2021). IEEE

    2021

  27. [35]

    Science 269(5232), 1880–1882 (1995)

    Wolpert, D.M., Ghahramani, Z., Jordan, M.I.: An internal model for sensorimotor integration. Science 269(5232), 1880–1882 (1995)

    1995

  28. [36]

    Land, M.F., Hayhoe, M.: In what ways do eye movements contribute to everyday activities? Vision research 41(25-26), 3559–3565 (2001)

    2001

  29. [37]

    Journal of neuroscience 21(17), 6917–6932 (2001)

    Johansson, R.S., Westling, G., B¨ ackstr¨ om, A., Flanagan, J.R.: Eye–hand coordination in object manipulation. Journal of neuroscience 21(17), 6917–6932 (2001)

    2001

  30. [38]

    Journal of Experimental Psychology: Human perception and perfor- mance 24(3), 830 (1998)

    Tucker, M., Ellis, R.: On the relations between seen objects and components of potential actions. Journal of Experimental Psychology: Human perception and perfor- mance 24(3), 830 (1998)

    1998

  31. [39]

    Human Movement Science 18(2-3), 201–218 (1999)

    Jeannerod, M.: Visuomotor channels: Their integration in goal-directed prehension. Human Movement Science 18(2-3), 201–218 (1999)

    1999

  32. [40]

    Pattern recognition 106, 107404 (2020)

    Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O.R., Jagersand, M.: U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition 106, 107404 (2020)

    2020

  33. [41]

    arXiv preprint arXiv:2306.12156 (2023)

    Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., 21 Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)

    2023 arXiv

  34. [42]

    In: Data Min- ing Workshops (ICDMW), 2017 IEEE Inter- national Conference On, pp

    McInnes, L., Healy, J.: Accelerated hierarchi- cal density based clustering. In: Data Min- ing Workshops (ICDMW), 2017 IEEE Inter- national Conference On, pp. 33–42 (2017). IEEE

    2017

  35. [43]

    arXiv preprint arXiv:2305.17110 (2023)

    Tang, B., Lin, M.A., Akinola, I., Handa, A., Sukhatme, G.S., Ramos, F., Fox, D., Narang, Y.: Industreal: Transferring contact- rich assembly tasks from simulation to reality. arXiv preprint arXiv:2305.17110 (2023)

    2023 arXiv

  36. [44]

    In: Euro- pean Robotics Forum, pp

    Albardaner, J., Miguel, A.S., Garc ´ ıa, N., Dal- mau, M.: Sim-to-real gap in rl: Use case with tiago and isaac sim/gym. In: Euro- pean Robotics Forum, pp. 344–348 (2024). Springer

    2024

  37. [45]

    arXiv preprint arXiv:2501.02902 (2025)

    Salimpour, S., Pe˜ na-Queralta, J., Paez- Granados, D., Heikkonen, J., Westerlund, T.: Sim-to-real transfer for mobile robots with reinforcement learning: from nvidia isaac sim to gazebo and real ros 2 robots. arXiv preprint arXiv:2501.02902 (2025)

    2025 arXiv

  38. [46]

    Evolving Systems 12(1), 217– 223 (2021)

    Victoria, A.H., Maragatham, G.: Automatic tuning of hyperparameters using bayesian optimization. Evolving Systems 12(1), 217– 223 (2021)

    2021

  39. [47]

    In: 2019 Third IEEE International Conference on Robotic Computing (IRC), pp

    Sehgal, A., La, H., Louis, S., Nguyen, H.: Deep reinforcement learning using genetic algorithm for parameter optimization. In: 2019 Third IEEE International Conference on Robotic Computing (IRC), pp. 596–601 (2019). IEEE

    2019

  40. [48]

    In: Conference on Robot Learning, pp

    Mandlekar, A., Nasiriany, S., Wen, B., Aki- nola, I., Narang, Y., Fan, L., Zhu, Y., Fox, D.: Mimicgen: A data generation system for scalable robot learning using human demon- strations. In: Conference on Robot Learning, pp. 1820–1864 (2023). PMLR

    2023

  41. [49]

    Batter- ies 10(5), 147 (2024)

    Rastegarpanah, A., Mineo, C., Contreras, C.A., Aflakian, A., Paragliola, G., Stolkin, R.: Electric vehicle battery disassembly using interfacing toolbox for robotic arms. Batter- ies 10(5), 147 (2024)

    2024

  42. [50]

    Memorie della Societa Astronomica Italiana 94(1), 114 (2023)

    Mineo, C., Maniscalco, U., Storniolo, P.: Edge robotics and emerging platforms with sensing and human interaction capabilities. Memorie della Societa Astronomica Italiana 94(1), 114 (2023)

    2023

  43. [51]

    22541/au.174533851.11035371/v1

    Rastegarpanah, A., Mineo, C., Contreras, C.A., Shaarawy, A., Paragliola, G., Stolkin, R.: Haptic teleoperation in extended real- ity for ev battery disassembly using gaussian mixture regression (2025) https://doi.org/10. 22541/au.174533851.11035371/v1

    2025

  44. [52]

    In: Proceedings of the 2024 4th International Conference on Robotics and Control Engineering, pp

    Rastegarpanah, A., Contreras, C.A., Stolkin, R.: Semi-autonomous robotic disassembly enhanced by mixed reality. In: Proceedings of the 2024 4th International Conference on Robotics and Control Engineering, pp. 7–13 (2024)

    2024

  45. [53]

    In: 2018 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pp

    Adjigble, M., Marturi, N., Ortenzi, V., Rajasekaran, V., Corke, P., Stolkin, R.: Model-free and learning-free grasping by local contact moment matching. In: 2018 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pp. 2933– 2940 (2018). IEEE

    2018

  46. [54]

    In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

    Adjigble, M., De Farias, C., Stolkin, R., Marturi, N.: Spectgrasp: Robotic grasping by spectral correlation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3987–3994 (2021). IEEE

    2021

  47. [55]

    arXiv preprint arXiv:2503.19786 (2025)

    Team, G., Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., Perrin, S., Mate- jovicova, T., Ram´ e, A., Rivi` ere, M., et al.: Gemma 3 technical report. arXiv preprint arXiv:2503.19786 (2025)

    2025 arXiv

  48. [56]

    arXiv preprint arXiv:2503.20020 (2025) 22

    Team, G.R., Abeyruwan, S., Ainslie, J., Alayrac, J.-B., Arenas, M.G., Armstrong, T., Balakrishna, A., Baruch, R., Bauza, M., Blokzijl, M., et al.: Gemini robotics: Bring- ing ai into the physical world. arXiv preprint arXiv:2503.20020 (2025) 22

    2025 arXiv

Pith tools

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