Efficient Manipulation-Enhanced Semantic Mapping With Uncertainty-Informed Action Selection

Jesper M\"ucke; Maren Bennewitz; Nils Dengler; Rohit Menon

REVIEW 3 major objections 5 minor 26 references

Efficient Manipulation-Enhanced Semantic Mapping With Uncertainty-Informed Action Selection

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Using its own evidential map uncertainty to steer viewpoints and pushes, the proposed pipeline cuts per-action planning time from roughly 40 seconds to under 2 seconds on heavily occluded shelves while roughly halving object displacement…

desk verdict Real speed and safety gains in simulation, but the mapping-accuracy numbers are computed from the predictor's own belief, so the accuracy story is not yet demonstrated. read the letter →

arxiv 2506.02286 v2 pith:GCJCCU2M submitted 2025-06-02 cs.RO

Nils Dengler , Jesper Mücke , Rohit Menon , Maren Bennewitz This is my paper

classification cs.RO

keywords manipulation-enhancedmappingnext-bestviewplanninguncertainty-informedactionselectionevidentialsemanticreinforcementlearningpushinteractiveperceptionservicerobots

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 tries to establish that a service robot can map heavily cluttered shelves by letting its own measured uncertainty decide where to look and what to push. The map is evidential: each voxel's occupancy is treated as a Beta distribution, a probability with an explicit uncertainty, and each cell's semantic label as a Dirichlet distribution, a set of class probabilities with a confidence score. The paper uses those uncertainty values as the reward for a reinforcement-learned viewpoint policy and as the score for an uncertainty-informed push sampler. If the experiments are right, the combined pipeline plans each action in about 1.66 seconds instead of 40.09 seconds, needs 2.59 pushes per scene instead of 5.25, moves objects about 1.00 meter instead of 1.93 meters, and collides in 6.7% of actions instead of 25.3%. The trade-off is lower final occupancy and semantic accuracy (0.846 and 0.653 versus 0.877 and 0.720 on the overlap metric), which the authors present as an acceptable price for speed and gentle interaction.

What carries the argument

The load-bearing mechanism is the evidential metric-semantic belief map, in which each voxel's occupancy is a Beta distribution with predictive mean and epistemic variance $u_o = \mathrm{Var}[\mathrm{Beta}(\alpha,\beta)]$, and each top-down cell's semantic label is a Dirichlet distribution with uncertainty $u_s = N/S$. This single representation does three jobs: it is the observation and reward signal for the RL viewpoint policy, it supplies the distance-transform and corridor scores for the push sampler, and it lets both kinds of actions be compared under the same expected-information-gain criterion. Around this map, a learned manipulation predictor (the CNABU model carried over from the prior pipeline) takes a proposed push parameterization and predicts the post-push evidential map, so a push can be evaluated before execution exactly like a viewpoint.

What would settle it

A quantitative real-robot rollout of the same 25-scene protocol would settle it: record per-iteration planning time, push count, displacement, and collisions, and compare predicted post-push occupancy and semantic accuracy against actual sensor re-observations. If predicted maps after pushes are consistently worse than re-scanning, the belief-update shortcut fails; if planning time and collision statistics depart from the reported 1.66 seconds and 6.7% beyond one standard deviation, the claimed zero-shot transfer is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that evidential uncertainty can act as one coherent steering signal for the whole interactive-perception loop. The same $u_o = \mathrm{Var}[\mathrm{Beta}(\alpha,\beta)]$ occupancy variance and $u_s = N/S$ semantic uncertainty that expose weak regions of the map feed three components: a soft-actor-critic reinforcement-learning policy that outputs continuous SE(3) camera poses in a single forward pass, a push sampler that extracts occlusion-critical corridors from an uncertainty-weighted distance transform and samples minimally invasive push directions from low-occupancy sectors, and an action selector that compares expected occlusion-aware information gain of the best view against simulated pushes. The paper reports that replacing the prior system's discrete viewpoint search and random push sampling with these uncertainty-guided components preserves most of the map quality while dramatically reducing planning cost, and that semantic uncertainty works better than occupancy uncertainty alone for selecting what to push.

Load-bearing premise

The pipeline assumes that the learned push simulator's predicted post-push map is accurate enough to be trusted as the new map after each real push, and that this transfers to the physical robot; the only real-robot validation is a qualitative video rather than a quantitative test.

Editorial extensions

If this is right

If the reported timings hold, manipulation-enhanced mapping becomes fast enough for online use: a per-decision budget under two seconds instead of forty.
The reduced push count and displacement mean the robot disturbs the shelf less, making the approach safer for real household objects and for objects that should stay put.
Because the viewpoint policy is trained only on uncertainty rewards, it removes the need for privileged ground-truth feedback at deployment, as the privileged-reward ablation supports.
Semantic uncertainty as the targeting signal should generalize to other occlusion-heavy scenes, since it directs the robot to regions where class labels are unresolved rather than merely unobserved geometry.
Replacing random push sampling with uncertainty-informed corridors improves or maintains map quality while cutting collisions and displacement, so the sampling strategy, not the underlying predictor alone, is responsible for much of the safety gain.

Reading between the lines

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

The same sensing-versus-acting arbitration could be ported to other confined-space tasks, such as inspection, assembly, fruit mapping, or shelf retrieval, wherever a learned forward model can predict the effect of a manipulation on the belief map; the paper's mechanism does not depend on shelves specifically.
A testable refinement would be learning the 0.1 semantic-uncertainty threshold that separates certain from uncertain cells, since the paper fixes it empirically; adapting it per scene could shift the mIoU-versus-push-count trade-off.
The real-robot evidence is only a qualitative zero-shot video, so a quantitative real-hardware sweep over the same 25-scene protocol is the natural next step and would either strengthen or undermine the claimed transfer.
Because the unmodified POMDP-push baseline still ends with higher mIoU, a hybrid policy that uses the fast RL viewpoints but falls back to the more thorough push evaluation on very cluttered scenes could recover some of the average accuracy without giving up all of the speed gain.

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.

Desk Editor's Note

read the letter

Two things to know before you read this. The efficiency and safety results are real, measured improvements: the RL next-best-view policy cuts planning time dramatically, and the uncertainty-informed push sampler cuts mean displacement, push count, and collision rate relative to random push sampling in the same simulator. The map accuracy claims are the soft spot. After the robot executes a push, the system takes the manipulation-CNABU's predicted belief as the map representation and never fuses a new observation, and the reported mIoU is computed from that same network output. That design means the semantic and occupancy numbers can be high even if the predictor is wrong, as long as it is confidently wrong. The stress-test note is on target here.

What is actually new: a SAC-trained continuous next-best-view policy operating on evidential maps, and a corridor-based push sampler that scores occlusion and free-space direction. The POMDP-CNABU-Informed ablation is the cleanest evidence that the push sampler does real work: it improves over random pushes on displacement, pushes per scene, and collisions while roughly matching mIoU. The RL policy's 99.7% inference-time reduction over the POMDP view planner is also well supported by Table III. Credit where due: this is the first version of this line of work that looks fast enough for online use.

Now the soft spots, in proportion. The prediction-as-map issue is the big one. It is load-bearing for every semantic and occupancy accuracy claim, and the only physical validation is a qualitative zero-shot video. No quantitative check of predicted versus realized post-push belief. I would not publish the accuracy claim as it stands. Second, the baseline protocol was changed: the completion threshold was raised from 95% to 99% for the reimplemented POMDP-CNABU-Push, which increases its push count and displacement. That does not invalidate the timing or ablation results, but the headline comparison against [10] should report this as a modified baseline. Third, three trials with no statistical tests is thin, though the effect sizes are large enough that this is a minor issue.

Circularity from using the same CNABU uncertainty for target selection, push scoring, and map update is real but not decisive, since the efficiency metrics are measured independently.

Who this is for: people working on active perception and interactive mapping in clutter. I would bring it to a reading group and I would send it to review, but with a requested revision: after a push, actually observe the scene again, or at least validate predicted beliefs against stored observations. Without that, the paper's contribution should be narrowed to planning efficiency and push safety in simulation. If the authors can add that validation, this becomes a solid systems paper.

Referee Report

3 major / 5 minor

Summary. The paper proposes a manipulation-enhanced semantic mapping framework that combines evidential metric-semantic belief maps (CNABU) with a Soft Actor-Critic continuous next-best-view planner and an uncertainty-informed push sampler. In simulation, the full system is reported to reduce planning time by ~95% (1.66 s vs 40.09 s per iteration), reduce the number of pushes (2.59 vs 5.25 per scene), lower mean object displacement (0.995 m vs 1.927 m), and lower collision rate (6.67% vs 25.33%) compared with the reimplemented POMDP-CNABU-Push baseline, at the cost of lower occupancy and semantic mIoU (0.846/0.653 vs 0.877/0.720). The real-hardware validation is a qualitative zero-shot video. The authors also ablate their push sampler by plugging it into the prior POMDP planner and confirm improved efficiency and reduced collisions.

Significance. If the experimental claims hold, the paper makes a useful practical contribution: the RL-based continuous view planner reduces planning cost dramatically, and the informed push sampler appears to select more targeted, minimally invasive actions than random push sampling. The ablations (POMDP-CNABU-Informed versus POMDP-CNABU-Push) are well designed and isolate the effect of the push-selection mechanism. However, the central mapping-accuracy claim is currently supported only through the CNABU network's own predicted beliefs, which is a self-referential evaluation. The headline efficiency numbers (planning time, displacement, collisions) are directly measured and independent of that concern, but the 'accurate mapping' claim needs additional validation before the paper can be accepted.

major comments (3)

[Section III-D and Section IV-B] The reported mIoU is computed on the CNABU network output after transformation into the map representation (Section IV-B), and after a selected push the system 'takes its predicted belief generated by the manipulation-CNABU as map representation' (Section III-D) without fusing any post-push sensor observation. This makes the occupancy and semantic mIoU curves, as well as the information-gain scores used for action selection, measures of the predictor's self-consistency rather than of the robot's actual map accuracy in the simulated environment. The claim that the framework 'enables a robot to construct complete semantic maps in cluttered, occlusion-heavy shelf environments' is therefore not established by the current experiments. Please either fuse actual post-push observations in simulation before computing metrics, or add a dedicated experiment comparing predicted post-push beliefs against the actual simulated post-push state, or explicitly and consistently reframe the claims as being about predicted-map quality and validate that predictor separately.
[Section III-C.4 and Section IV-C.2] The manipulation-CNABU predictor was trained in prior work [10] on its own push parameterizations. The new corridor-based push sampling in Section III-C.4 generates pushes with directions and contact points that may be out of distribution for that predictor. No quantitative comparison is provided between the CNABU-predicted post-push belief and the actual post-push state (e.g., from the PyBullet simulator's ground truth) for these newly generated pushes. This is load-bearing because the VIG push score, the decision to push, and the resulting map representation all flow through this forward model. A bias in the predictor for the new pushes would compound over the 40-step budget and could explain the mIoU differences. Please add a direct predicted-versus-actual post-push evaluation for the corridor-sampled pushes.
[Section IV-C] The description of the changed completion threshold is ambiguous and potentially unfair to the baseline. The text states: 'However, we increased the certainty threshold for full map completion after which no pushes are performed from 95%, as stated in [10], to 99%.' It is unclear whether this change was applied only to the proposed method or also to the reimplemented POMDP-CNABU-Push baseline. If applied only to the proposed method, the comparison is not controlled; if applied to all methods, the baseline no longer matches the published configuration of [10] and the effect of this change should be analyzed explicitly. Please clarify and, if necessary, report results under the original 95% threshold as well.

minor comments (5)

[Section IV-C.2] The statement that 'the worst performing interactive perception method achieves higher semantic mIOU than the best active perception method' is contradicted by Table IV: POMDP-CNABU-Push (ToF), an interactive perception variant, has semantic mIoU 0.638, while the best active method POMDP-CNABU-View has semantic mIoU 0.641.
[Table IV caption] The caption mentions the variant 'us|ub', but the table rows list 'us|uo' and 'uo|uo'; the caption should be corrected to match the actual configuration names.
[Section III-C.1] The semantic uncertainty threshold of 0.1 and the empirical weights k1-k4 in Eq. (5) are stated to be chosen empirically, but no sensitivity analysis or selection procedure is described; please provide at least a brief justification or a reference to a supplementary study.
[Section IV-B] The mIoU metric definition says the IoU is 'calculated on the CNABU network output after transformation into the map representation'; this conflation of prediction and robot map is the source of the major concern above and should be clarified in the main text so that readers understand what is being measured.
[Abstract] Minor language issues: 'enables to accurately map' should be 'enables the robot to accurately map' or similar, and the Figure 1 caption contains an incomplete phrase ('informed push sampler then most informative uncertain point').

Circularity Check

2 steps flagged · score 4.0 of 10

Timing, displacement, and collision claims are independently grounded in simulation, but the interactive map-accuracy verdict is partly self-referential: after each push the manipulation-CNABU's own predicted belief is installed as the map state, and the reported mIoU is computed on that same CNABU output, with no post-push observation ever fused.

self definitional [Section III-D (Manipulation-Enhanced Planning); Section IV-B (Metrics)]
"Otherwise, if pushing is expected to yield higher information gain, the robot proceeds to execute the selected push to actively modify the scene and takes its predicted belief generated by the manipulation-CNABU as map representation. ... mIOU: Semantic and occupancy mean Intersection-over-Union, computed as the average per-class IoU between predictions and ground truth, calculated on the CNABU network output after transformation into the map representation."

A push is selected because the manipulation-CNABU predicts a high VIGpush from its own post-push evidential map, and then that same predicted belief is adopted as the actual map state without fusing any post-push RGB-D observation. The expected information gain is therefore realized by construction: the belief becomes whatever the predictor asserted. The next VIG computation and the headline mIoU metric both operate on this same CNABU output, so the pipeline's map-accuracy claim cannot distinguish genuinely revealed regions from a confidently wrong open-loop prediction.
self citation load bearing [Section III-A; Section IV-C (reimplementation note)]
"Following the CNABU framework [10], we represent the map belief as a probabilistic estimate over occupancy and semantic categories, enriched with explicit uncertainty. ... To compare our approach against the POMDP-CNABU planners [10], we re-implemented their pipeline and used the provided CNABU-models from their code release. In order to use the models for our approach, no fine-tuning on the novel set of continuous viewpoints was performed."

The load-bearing premise that CNABU produces accurate post-push and post-view beliefs for the new continuous viewpoints and corridor-sampled pushes is imported from [10], whose author list (Marques, Dengler, Muecke, Bennewitz) overlaps the present paper. The models are used as released without fine-tuning, and the only physical validation is a qualitative zero-shot hardware video, so the accuracy of predicted versus realized post-push beliefs for exactly these pushes is assumed rather than demonstrated. Mitigating factors: [10] is a published, code-released, peer-reviewed result with its own external evaluations, and the present paper compares against a reimplemented version of that same pipeline, so the citation is real evidence rather than a closed loop.

full rationale

The majority of the paper's headline claims are not circular. Planning-time reductions (Table III) are pure computational measurements, and mean displacement, push counts, and collision rates are measured in the PyBullet physics simulator against the actual executed pushes, independent of any Beta/Dirichlet uncertainty value. The experiments include randomized (Random-View) and privileged (RL-CNABU-View-Privileged, ground-truth IoU reward) baselines, and the final mIoU is computed against ground truth, so the action-selection strategies are externally benchmarked rather than merely internally consistent. The circularity is concentrated in the interactive mapping loop: Section III-D installs the manipulation-CNABU's predicted post-push belief as the map representation with no post-push sensor observation, and Section IV-B computes the reported semantic and occupancy mIoU on that same CNABU output. Consequently, a push's predicted information gain is realized by construction in the belief space, and the map-quality verdict partly measures the predictor's open-loop self-consistency for the new corridor-sampled pushes rather than a sensor-grounded robot map. This is a genuine partial self-definitional issue, but it does not infect the efficiency, displacement, or collision claims, and the citations to [10] carry real code-released evidence. The honest finding is therefore a moderate circularity score of 4, not higher: some of the mapping-accuracy derivation reduces to the predictor's own output, while the central efficiency contribution has independent content.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the CNABU belief predictor from the same research group, several hand-tuned thresholds and weights, and the assumption that simulation transfers to reality. No new physical entities are introduced. The independent evidence comes from ground-truth mIoU and from random and privileged baselines within the paper.

free parameters (8)

semantic uncertainty threshold = 0.1
Cells with semantic uncertainty below 0.1 are treated as certain and set to zero during target identification (Section III-C.1); the value is empirically chosen.
corridor scoring weights k1, k2, k3, k4 = 2, 3, 4, 5
Weights in Eq. (5) for scoring visibility corridors; described as empirically found in Table II.
max candidate target locations = 5
A maximum of five cells with highest distance values are selected as push targets (Section III-C.1).
pushing corridor sector angle = 30 degrees
Rays are grouped into sectors with maximum angle 30 degrees for push direction evaluation (Section III-C.3).
reward weights w1 and w2 = 10, 2
Weights in the RL reward Eq. (2); listed in Table II.
repeat penalty thresholds theta_p and theta_e = 0.1 m, 0.034 rad
Distance thresholds in the repetition penalty Eq. (3); listed in Table II.
map completion certainty threshold = 99%
Raised from 95% in [10] to 99% to encourage further pushes during evaluation (Section IV-C).
corridor radius n_c and push candidate count n_p
Used for ray casting and push sampling in Section III-C; exact values are not reported, hampering exact reproduction.

assumptions (4)

domain assumption The CNABU evidential predictor supplies calibrated occupancy and semantic uncertainty and accurate post-manipulation belief maps.
Section III-A and III-D use these predicted beliefs as the map representation and as the basis for both reward and push scoring; no online correction against real observations is provided.
domain assumption PyBullet physics and the modeled Realsense L515 camera are faithful enough for training and evaluation to transfer to real hardware.
All quantitative results come from PyBullet (Section IV-A); real-robot evidence is a qualitative zero-shot video (Section IV-E).
domain assumption The SAC policy trained on 12,500 procedurally generated shelf configurations generalizes to the 25 evaluation scenes and to real scenes.
Section IV-A and IV-C evaluate without fine-tuning; this assumption is load-bearing for the reported planning speed and mapping behavior.
ad hoc to paper Uncertainty from Beta and Dirichlet distributions is an adequate proxy for expected information gain in view and push selection.
This postulate is the paper's design choice (Sections III-B.3 and III-C.1). It is tested only indirectly through the ablations and comparisons to random baselines.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient Manipulation-Enhanced Semantic Mapping With Uncertainty-Informed Action Selection." pith.science (2026). https://pith.science/paper/GCJCCU2M

@misc{pith2026250602286,
  author       = {Pith},
  title        = {Pith review of: Efficient Manipulation-Enhanced Semantic Mapping With Uncertainty-Informed Action Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GCJCCU2M}},
  note         = {Machine review of arXiv:2506.02286}
}

read the original abstract

Service robots operating in cluttered human environments such as homes, offices, and schools cannot rely on predefined object arrangements and must continuously update their semantic and spatial estimates while dealing with possible frequent rearrangements. Efficient and accurate mapping under such conditions demands selecting informative viewpoints and targeted manipulations to reduce occlusions and uncertainty. In this work, we present a manipulation-enhanced semantic mapping framework for occlusion-heavy shelf scenes that integrates evidential metric-semantic mapping with reinforcement-learning-based next-best view planning and targeted action selection. Our method thereby exploits uncertainty estimates from Dirichlet and Beta distributions in the map prediction networks to guide both active sensor placement and object manipulation, focusing on areas with high uncertainty and selecting actions with high expected information gain. Furthermore, we introduce an uncertainty-informed push strategy that targets occlusion-critical objects and generates minimally invasive actions to reveal hidden regions by reducing overall uncertainty in the scene. The experimental evaluation shows that our framework enables to accurately map cluttered scenes, while substantially reducing object displacement and achieving a 95% reduction in planning time compared to the state-of-the-art, thereby realizing real-world applicability.

Figures

Figures reproduced from arXiv: 2506.02286 by the authors.

**Figure 2.** Overview of our uncertainty-aware manipulation-enhanced mapping framework. Our system integrates evidential semantic mapping [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Uncertainty-informed push sampling. From the belief map [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Simulation results across 25 highly cluttered shelf environments with an action budget of 40 steps for RL-CNABU-Push and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

[10]

Map Space Belief Prediction for Manipulation-Enhanced Mapping,

J. M. C. Marques, N. Dengler, T. Zaenker, J. Muecke, W. Shenlong, M. Bennewitz, and K. Hauser, “Map Space Belief Prediction for Manipulation-Enhanced Mapping,” in Proc. of Robotics: Science and Systems (RSS), 2025

work page 2025
[1]

Computer Vision for 3D Perception: A Review,

N. O’Mahony, S. Campbell, L. Krpalkova, D. Riordan, J. Walsh, A. Murphy, and C. Ryan, “Computer Vision for 3D Perception: A Review,” in Proc. of the Intelligent Systems Conference (IntelliSys) , 2019

work page 2019
[2]

Roboexp: Action-Conditioned Scene Graph via Interactive Exploration for Robotic Manipulation,

H. Jiang, B. Huang, R. Wu, Z. Li, S. Garg, H. Nayyeri, S. Wang, and Y . Li, “Roboexp: Action-Conditioned Scene Graph via Interactive Exploration for Robotic Manipulation,” Proc. of Conf. on Robot Learning (CoRL), 2024

work page 2024
[3]

Mechanical Search on Shelves With Efficient Stacking and Destacking of Objects,

H. Huang, L. Fu, M. Danielczuk, C. M. Kim, Z. Tam, J. Ichnowski, A. Angelova, B. Ichter, and K. Goldberg, “Mechanical Search on Shelves With Efficient Stacking and Destacking of Objects,” in Proc. of the Intl. Symposium on Robotic Research (ISRR) . Springer, 2022

work page 2022
[4]

Viewpoint Push Planning for Mapping of Unknown Confined Spaces,

N. Dengler, S. Pan, V . Kalagaturu, R. Menon, M. Dawood, and M. Bennewitz, “Viewpoint Push Planning for Mapping of Unknown Confined Spaces,” in Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2023

work page 2023
[5]

SpotLight: Robotic Scene Understanding through Interaction and Affordance Detection,

T. Engelbracht, R. Zurbr ¨ugg, M. Pollefeys, H. Blum, and Z. Bauer, “SpotLight: Robotic Scene Understanding through Interaction and Affordance Detection,” arXiv preprint arXiv:2409.11870 , 2024

arXiv 2024
[6]

Safe Leaf Manipulation for Accurate Shape and Pose Estimation of Occluded Fruits,

S. Yao, S. Pan, M. Bennewitz, and K. Hauser, “Safe Leaf Manipulation for Accurate Shape and Pose Estimation of Occluded Fruits,” in Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2025

work page 2025
[7]

Safe, occlusion-aware manipulation for online object reconstruction in confined spaces,

Y . Miao, R. Wang, and K. Bekris, “Safe, occlusion-aware manipulation for online object reconstruction in confined spaces,” in Proc. of the Intl. Symposium on Robotic Research (ISRR) . Springer, 2022

work page 2022

Show all 26 references

[8]

Selective object rearrangement in clutter,

B. Tang and G. S. Sukhatme, “Selective object rearrangement in clutter,” in Proc. of the Conference on Robot Learning. PMLR, 2023

2023
[9]

Visual Foresight Trees for Object Retrieval from Clutter with Nonprehensile Rearrangement,

B. Huang, S. D. Han, J. Yu, and A. Boularias, “Visual Foresight Trees for Object Retrieval from Clutter with Nonprehensile Rearrangement,” IEEE Robotics and Automation Letters (RA-L) , 2021

2021
[11]

Active Visuo-Tactile Interactive Robotic Perception for Accurate Object Pose Estimation in Dense Clutter,

P. K. Murali, A. Dutta, M. Gentner, E. Burdet, R. Dahiya, and M. Kaboli, “Active Visuo-Tactile Interactive Robotic Perception for Accurate Object Pose Estimation in Dense Clutter,” IEEE Robotics and Automation Letters (RA-L) , 2022

2022
[12]

ActNeRF: Uncertainty- Aware Active Learning of NeRF-Based Object Models for Robot Manipulators using Visual and Re-orientation Actions,

S. Dasgupta, A. Gupta, S. Tuli, and R. Paul, “ActNeRF: Uncertainty- Aware Active Learning of NeRF-Based Object Models for Robot Manipulators using Visual and Re-orientation Actions,” in Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2024

2024
[13]

Reactive Informative Planning for Mobile Manipulation Tasks Under Sensing and Environmental Uncertainty,

M. Tzes, V . Vasilopoulos, Y . Kantaros, and G. J. Pappas, “Reactive Informative Planning for Mobile Manipulation Tasks Under Sensing and Environmental Uncertainty,” in Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2022

2022
[14]

Uncertainty-Aware Manipulation Plan- ning Using Gravity and Environment Geometry,

F. von Drigalski, K. Kasaura, C. C. Beltran-Hernandez, M. Hamaya, K. Tanaka, and T. Matsubara, “Uncertainty-Aware Manipulation Plan- ning Using Gravity and Environment Geometry,” IEEE Robotics and Automation Letters (RA-L) , 2022

2022
[15]

Evidential Deep Learning to Quantify Classification Uncertainty,

M. Sensoy, L. Kaplan, and M. Kandemir, “Evidential Deep Learning to Quantify Classification Uncertainty,” Advances in Neural Information Processing Systems, 2018

2018
[16]

vMF-Contact: Uncertainty-aware Evidential Learning for Probabilistic Contact-grasp in Noisy Clutter,

Y . Shi, E. Welte, M. Gilles, and R. Rayyes, “vMF-Contact: Uncertainty-aware Evidential Learning for Probabilistic Contact-grasp in Noisy Clutter,” Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2025

2025
[17]

Uncertainty Estimation for 3D Object Detection via Evidential Learning,

N. Durasov, R. Mahmood, J. Choi, M. T. Law, J. Lucas, P. Fua, and J. M. Alvarez, “Uncertainty Estimation for 3D Object Detection via Evidential Learning,” arXiv preprint arXiv:2410.23910 , 2024

2024 arXiv
[18]

Semantic Mechanical Search with Large Vision and Language Models,

S. Sharma, H. Huang, K. Shivakumar, L. Y . Chen, R. Hoque, brian ichter, and K. Goldberg, “Semantic Mechanical Search with Large Vision and Language Models,” in Proc. of Conf. on Robot Learning (CoRL), 2023

2023
[19]

Mechanical Search: Multi-Step Retrieval of a Target Object Occluded by Clutter,

M. Danielczuk, A. Kurenkov, A. Balakrishna, M. Matl, D. Wang, R. Mart´ın-Mart´ın, A. Garg, S. Savarese, and K. Goldberg, “Mechanical Search: Multi-Step Retrieval of a Target Object Occluded by Clutter,” in Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2019

2019
[20]

Curi- ousBot: Interactive Mobile Exploration via Actionable 3D Relational Object Graph,

Y . Wang, L. Fermoselle, T. Kelestemur, J. Wang, and Y . Li, “Curi- ousBot: Interactive Mobile Exploration via Actionable 3D Relational Object Graph,” arXiv preprint arXiv:2501.13338 , 2025

2025
[21]

Robust Contact Generation for Robot Simulation with Unstructured Meshes,

K. Hauser, “Robust Contact Generation for Robot Simulation with Unstructured Meshes,” in Proc. of the Intl. Symposium on Robotic Research (ISRR), 2013

2013
[22]

NBV-SC: Next Best View Planning Based on Shape Completion for Fruit Mapping and Reconstruction,

R. Menon, T. Zaenker, N. Dengler, and M. Bennewitz, “NBV-SC: Next Best View Planning Based on Shape Completion for Fruit Mapping and Reconstruction,” in Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2023

2023
[23]

Occlusion- Aware Search for Object Retrieval in Clutter,

W. Bejjani, W. C. Agboh, M. R. Dogar, and M. Leonetti, “Occlusion- Aware Search for Object Retrieval in Clutter,” inProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2021

2021
[24]

PyBullet, a Python Module for Physics Simu- lation for Games, Robotics and Machine Learning,

E. Coumans and Y . Bai, “PyBullet, a Python Module for Physics Simu- lation for Games, Robotics and Machine Learning,” http://pybullet.org
[25]

Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning With a Stochastic Actor,

T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning With a Stochastic Actor,” in Proc. of the Intl. Conf. on Machine Learning , 2018

2018
[26]

Stable-Baselines3: Reliable Reinforcement Learning Implementations,

A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-Baselines3: Reliable Reinforcement Learning Implementations,” Journal of Machine Learning Research , 2021

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [10]

Map Space Belief Prediction for Manipulation-Enhanced Mapping,

J. M. C. Marques, N. Dengler, T. Zaenker, J. Muecke, W. Shenlong, M. Bennewitz, and K. Hauser, “Map Space Belief Prediction for Manipulation-Enhanced Mapping,” in Proc. of Robotics: Science and Systems (RSS), 2025

work page 2025

[2] [1]

Computer Vision for 3D Perception: A Review,

N. O’Mahony, S. Campbell, L. Krpalkova, D. Riordan, J. Walsh, A. Murphy, and C. Ryan, “Computer Vision for 3D Perception: A Review,” in Proc. of the Intelligent Systems Conference (IntelliSys) , 2019

work page 2019

[3] [2]

Roboexp: Action-Conditioned Scene Graph via Interactive Exploration for Robotic Manipulation,

H. Jiang, B. Huang, R. Wu, Z. Li, S. Garg, H. Nayyeri, S. Wang, and Y . Li, “Roboexp: Action-Conditioned Scene Graph via Interactive Exploration for Robotic Manipulation,” Proc. of Conf. on Robot Learning (CoRL), 2024

work page 2024

[4] [3]

Mechanical Search on Shelves With Efficient Stacking and Destacking of Objects,

H. Huang, L. Fu, M. Danielczuk, C. M. Kim, Z. Tam, J. Ichnowski, A. Angelova, B. Ichter, and K. Goldberg, “Mechanical Search on Shelves With Efficient Stacking and Destacking of Objects,” in Proc. of the Intl. Symposium on Robotic Research (ISRR) . Springer, 2022

work page 2022

[5] [4]

Viewpoint Push Planning for Mapping of Unknown Confined Spaces,

N. Dengler, S. Pan, V . Kalagaturu, R. Menon, M. Dawood, and M. Bennewitz, “Viewpoint Push Planning for Mapping of Unknown Confined Spaces,” in Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2023

work page 2023

[6] [5]

SpotLight: Robotic Scene Understanding through Interaction and Affordance Detection,

T. Engelbracht, R. Zurbr ¨ugg, M. Pollefeys, H. Blum, and Z. Bauer, “SpotLight: Robotic Scene Understanding through Interaction and Affordance Detection,” arXiv preprint arXiv:2409.11870 , 2024

arXiv 2024

[7] [6]

Safe Leaf Manipulation for Accurate Shape and Pose Estimation of Occluded Fruits,

S. Yao, S. Pan, M. Bennewitz, and K. Hauser, “Safe Leaf Manipulation for Accurate Shape and Pose Estimation of Occluded Fruits,” in Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2025

work page 2025

[8] [7]

Safe, occlusion-aware manipulation for online object reconstruction in confined spaces,

Y . Miao, R. Wang, and K. Bekris, “Safe, occlusion-aware manipulation for online object reconstruction in confined spaces,” in Proc. of the Intl. Symposium on Robotic Research (ISRR) . Springer, 2022

work page 2022