Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

ScanBot: A Benchmark for Precision Robotic Surface Scanning with Industrial Laser Profilers

T0 review · 3 major / 3 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 ScanBot, a new benchmark, shows that today's AI models cannot perform precision laser scanning.

desk verdict A genuinely new dataset for instruction-conditioned laser scanning, but the central claim that current models cannot scan is not actually tested by the paper's own evaluation. read the letter →

arxiv 2505.17295 v2 pith:HDA7ZPTX submitted 2025-05-22 cs.RO

Zhiling Chen , Yang Zhang , Fardin Jalil Piran , Qianyu Zhou , Jiong Tang , Farhad Imani This is my paper · ORCID
classification cs.RO
keywords instruction-conditionedscanninglaserprofilerbenchmarkrobotlearningdatasetmultimodallargelanguagemodelssensor-as-actuatorsurfacereconstructionnon-destructiveevaluationembodiedAI
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

ScanBot is presented as the first benchmark built for instruction-conditioned surface scanning with an industrial laser profiler, where sub-millimeter path continuity and stable sensor settings decide whether usable geometry is acquired. The paper argues that this sensing-centric task is fundamentally different from gripper-based manipulation and is missing from existing robot-learning datasets. In its evaluation, four state-of-the-art multimodal language models fail across the perception–planning–execution loop: none reaches even 42% accuracy at choosing scanner parameters, mean localization IoU tops out at 0.129 across tasks and drops toward zero on fine-grained targets, and predicted scan paths land off the object footprint. If ScanBot's framing is right, robot learning needs new data, supervision, and evaluation for tools whose output is measured by measurement quality rather than task completion.

What carries the argument

The load-bearing object is the ScanBot dataset itself: 896 scripted scanning trajectories across six real electronic components and six 3D-printed parts, each paired with a natural-language instruction, synchronized first-person RGB-D and third-person video, laser height profiles, joint and pose traces, and scanner-parameter logs. The evaluation machinery that carries the argument is the two-stage pipeline that isolates the two skills the authors say industrial scanning requires: 'set up the sensor' (predict six discrete profiler parameters) and 'move like a scanner' (localize the region, then output start and end end-effector poses). A secondary mechanical choice is to simplify trajectory prediction to a start and end pose on flat surfaces, which makes the benchmark tractable while keeping the failure modes visible.

What would settle it

Run the same benchmark in a version that lets models emit a dense sequence of waypoints and correct course from live laser profiles while scanning a curved object; if any current model keeps the surface inside the profiler's ±34 mm range, holds trajectory jitter below the sensor's 0.1 mm depth resolution, and reconstructs the target region with Chamfer distance comparable to the expert scan, the paper's claim that learning-based models cannot produce stable precision scan motions would be overturned.

Watch

Extended reading notes

Core claim

The central claim is that current learning-based embodied models, including state-of-the-art multimodal large language models such as GPT-4.1, OpenAI o3, Gemini 2.5 Pro, and Gemini 2.5 Flash, cannot yet turn natural-language scan instructions into stable, feasible laser-profiling motions under real industrial constraints. The paper substantiates this through ScanBot, a dataset of 896 scanning trajectories over twelve objects and six task types, and through a two-stage evaluation: Stage I asks a model to set six scanner parameters from first- and third-person images, where the best average accuracy is 41.7%; Stage II asks it to ground an instruction to a 2D bounding box and then to output start and end end-effector poses, where average IoU is at most 0.129 and predicted waypoints consistently fall off the object footprint. Consequently, the acquired point clouds carry maximal reconstruction error rather than usable geometry. The paper takes this as evidence that scanner-based active perception, treating the sensor as the actuator, deserves its own benchmark and its own model design rather than being treated as another grasping task.

Load-bearing premise

The evaluation assumes that asking a model to draw a box around the target and name a start and end point is good enough to stand for the full, unbroken, sub-millimeter motion of a real scan; if that simplification misses the real challenge, the finding that models fail at precision scanning does not follow.

Editorial extensions

If this is right

  • A model that passes ScanBot's two stages would need both physical parameter reasoning and millimeter-scale language grounding; today's best MLLMs meet neither threshold.
  • Robot-learning benchmarks should add sensor-as-actuator tasks alongside manipulation, because coverage and measurement quality are the success criteria, not grasp success.
  • Industrial inspection of large parts could move from exhaustive full-surface scans to instruction-selected region scans if models can ground phrases like 'leading-edge weld' to exact surface regions.
  • The benchmark's failure profile identifies specific bottlenecks: Z-center and CMOS-range parameter prediction, and fine-grained localization on T2–T4 tasks, are where current models collapse.

Reading between the lines

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

  • If the start-end simplification understates the real control burden, ScanBot's negative results are optimistic; a version requiring full waypoint sequences with jitter limits near the sensor's 0.1 mm depth resolution could make the gap larger.
  • A natural next experiment, not run in the paper, is to split the pipeline: let a classical coverage planner handle waypoint generation and use an MLLM only for region grounding; ScanBot's annotations make this measurable and would reveal whether language grounding or motion control is the binding constraint.
  • The same two-stage 'set up the tool, move like the tool' design transfers to other precision tools the paper names—welders, sprayers, polishers—with their own tolerances such as a ±0.2 mm weld bead or a ±2 mm paint standoff.
  • Because surfaces in ScanBot are flat and scanning is open-loop, the benchmark does not expose closed-loop behavior; a plausible extension is a variant where models can replan from live laser profiles, which may be where learned policies ultimately surpass scripted planners.
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

3 major / 3 minor

Summary. The paper introduces ScanBot, presented as the first instruction-conditioned benchmark for robot-mounted industrial laser surface scanning. The dataset comprises 12 objects (six real electronic components and six 3D-printed shapes), 896 scanning paths across six task types, and synchronized first-person RGB-D, third-person video, laser height profiles, robot pose and joint traces, and scanner-parameter logs. To benchmark the dataset, the authors evaluate four MLLMs (GPT-4.1, OpenAI o3, Gemini 2.5 Pro, Gemini 2.5 Flash) on three proxy tasks: scanner-parameter selection, bounding-box region localization, and start/end pose prediction. They report low parameter-selection accuracy and low IoU, and conclude that current learning-based models fail to produce stable, feasible, precision scanning trajectories under fine-grained instructions. Section 5 acknowledges that all trajectories assume flat surfaces, are open-loop, and are single-pass.

Significance. If the headline claim were fully supported, ScanBot would fill a genuine gap: existing robot learning datasets focus on gripper-based manipulation, whereas active sensing with a laser profiler has different precision and control requirements. The dataset's synchronized multimodal recordings on real hardware, detailed scanner-parameter documentation, and explicit failure-case analysis (Appendix A.5) are useful contributions that could support future work on tool-conditioned perception and control. The paper is also transparent about its scope limitations. However, the evaluation as presented does not measure the central quantities that the title and abstract emphasize—motion continuity, standoff control, and trajectory feasibility—so the benchmark's evidentiary value for its main claim is currently limited. The mismatch between the VLA claim in the abstract and the MLLM-only experiments further weakens the significance of the reported negative results.

major comments (3)
  1. [§4.3, §6] The paper's central conclusion that current models fail to produce stable and feasible scan motions is not directly supported by the reported experiments. Section 4.3 reduces trajectory prediction to predicting only the start and end end-effector poses and says these are evaluated by Euclidean deviation and Chamfer distance, but no quantitative results, table, or error analysis for either metric are given; the only reported outcome is the qualitative statement that predicted way-points consistently fall off the object footprint. In addition, Section 5 concedes that all trajectories assume flat surfaces and are open-loop, so path continuity, jitter, standoff deviation, and coverage over a continuous trajectory are never assessed. Please either report direct trajectory-quality metrics on model-generated paths (for example, per-waypoint standoff error, jitter, coverage, and Chamfer distance to the reference scan) or explicitly limit the conclusion to the proxy tasks actually evaluated.
  2. [Abstract, §1, §4] The abstract and conclusion state that vision-language action (VLA) models fail at precision scanning, but the experiments evaluate four multimodal large language models (GPT-4.1, OpenAI o3, Gemini 2.5 Pro, Gemini 2.5 Flash) on static prediction tasks: parameter selection, bounding-box prediction, and endpoint pose prediction. None of these models is a vision-language-action policy, and no VLA model such as OpenVLA (Ref. [1]) or RT-1 (Ref. [22]) is tested. The claim that VLA models fail at continuous scanning is therefore an extrapolation from MLLM performance on simplified proxies. Please either evaluate actual VLA policies on the benchmark or rephrase the claim to state that current MLLMs fail on the evaluated proxy tasks.
  3. [Tables 3 and 4] The benchmark results are reported without statistical support. The accuracy values in Table 3 and the mean IoU values in Table 4 have no sample sizes, variance measures, confidence intervals, or significance tests. For example, the difference between Gemini 2.5 Flash at 41.7% and Gemini 2.5 Pro at 40.3% in Table 3 is presented as a ranking even though the values may be statistically indistinguishable. Since one purpose of the paper is to benchmark model capabilities, please include per-model and per-task trial counts and error bars or significance tests, and state how many instructions, objects, and runs each aggregate number represents.
minor comments (3)
  1. [Abstract] The abstract says the dataset contains scanning trajectories over twenty objects, but Section 3.2 and the caption of Figure 1 describe 12 objects (six real-world components and six 3D-printed shapes). Please correct the abstract to match the actual dataset size.
  2. [Table 1, §3.4] Table 1 lists ScanBot as having 197 tasks, while Section 3.4 reports 896 total scanning paths. Please clarify what the number 197 counts (for example, object-task combinations or episodes) and reconcile the two numbers so readers can correctly interpret the dataset size.
  3. [§4.3] Section 4.3 states that path quality is evaluated by Euclidean deviation and Chamfer distance, but no quantitative results for these metrics appear anywhere in the paper. If these metrics were computed, please include the numbers in the main text or an appendix; otherwise, remove the claim and keep only the qualitative failure description.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ScanBot's failure results are empirical benchmark outcomes, not self-referential predictions.

full rationale

The paper's central claim—that current MLLMs underperform on ScanBot's precision scanning tasks—is an empirical observation, not a derivation that reduces to its own inputs. The benchmark defines ground truth independently: scanner parameters were set by manual expert tuning during data collection (Sec. 3.3, App. A.2), and trajectories come from executed robot scans. Model performance is then measured on held-out prediction tasks (parameter selection accuracy in Sec. 4.1, IoU in Sec. 4.2, endpoint deviation and Chamfer distance in Sec. 4.3). None of these metrics is a renamed fit: the model outputs are compared with ground truth rather than used to define it. There are no self-citations carrying the argument, no imported uniqueness theorems, and no ansatz smuggled in via the authors' prior work. The main caveat is external validity, not circularity: Sec. 4.3 explicitly reduces trajectory generation to start/end pose prediction, so the conclusion about 'stable and feasible scan motions' is an extrapolation from proxy tasks. That is a fairness/generalization concern, not a case where an equation or parameter equals its own input by construction. The failure results could plausibly have come out differently, so they are genuine empirical findings. Section 5's limitations (flat surfaces, open-loop execution) further delimit the scope of the claims without creating circularity.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The paper introduces no new physical entities or free parameters in a model sense. The axioms are domain assumptions about the validity of expert-selected ground truth and the evaluation protocol. The dataset itself is the main contribution, but it is not yet accessible for independent checking.

assumptions (2)
  • domain assumption Expert-defined scanner parameters and trajectories are treated as the ground truth.
    The paper uses manual tuning during data collection as the reference for parameter prediction and trajectory evaluation, without a quantitative scan-quality metric to validate these choices (Sections 3.4 and 4.1).
  • domain assumption The simplified evaluation outputs (bounding box, start/end poses) are meaningful proxies for scanning performance.
    Section 4.3 explicitly simplifies trajectory prediction to start and end poses, and Section 4.2 uses 2D bounding boxes for localization. The paper assumes these capture the key difficulties of precision scanning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ScanBot: A Benchmark for Precision Robotic Surface Scanning with Industrial Laser Profilers." pith.science (2026). https://pith.science/paper/HDA7ZPTX

@misc{pith2026250517295,
  author       = {Pith},
  title        = {Pith review of: ScanBot: A Benchmark for Precision Robotic Surface Scanning with Industrial Laser Profilers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDA7ZPTX}},
  note         = {Machine review of arXiv:2505.17295}
}
read the original abstract

We introduce ScanBot, a benchmark for instruction-conditioned, high-precision surface scanning with robot-mounted industrial laser profilers. Unlike existing robot learning datasets that emphasize coarse behaviors such as grasping, navigation, or dialogue, ScanBot targets sensing-centric tasks where sub-millimeter motion continuity, strict stand-off control, and stable scanner settings are essential for acquiring usable geometry. The dataset contains scanning trajectories over twenty objects, including electronic components and structured 3D-printed parts, and spans six task types that range from broad inspection to fine-grained detail scanning and geometry-critical operations, including metrology and registration. Each episode is specified by natural language instructions and paired with synchronized first-person RGB-D, third-person video, laser height profiles, robot joint and pose traces, and scanner-parameter logs. These requirements expose a gap: despite recent progress, learning-based models often fail to produce stable and feasible scan motions under fine-grained instructions and real laser-profiling constraints. To reflect how industrial scanning is actually done, we evaluate methods through a two-stage pipeline. Stage I asks the model to "set up the sensor" by recommending scanner parameters, while Stage II asks it to "move like a scanner" by producing smooth, feasible trajectories that maintain stand-off and cover the intended region under precision demands.

Figures

Figures reproduced from arXiv: 2505.17295 by the authors.

Figure 1
Figure 1. Overview and motivation behind the ScanBot dataset. (a) Embodied AI must generalize not [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Hardware setup of the ScanBot system. A UR3 robotic arm is equipped with a Keyence [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Overview of the 12 scanned objects in the ScanBot dataset. The top two rows show six [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Multiview examples and annotated features in the ScanBot dataset. The first column shows [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Distribution of scanning tasks across six instruction [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Feature-level ground-truth scans for 3D-printed objects. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Feature-level ground-truth scans for real-world electronic components. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Measurement geometry of the Keyence LJ-X8200 scanner. At a reference distance of [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Scanner profile previews used for height calibration. (a) Profile centered at reference height. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Example of an instruction-conditioned scanning trajectory. The task “Scan the top surface [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Examples of failed surface reconstructions caused by inappropriate configurations. (a) A [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. How Should a Robot Configure Its Laser Scanner for Inspection?

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    SenseHD selects stable laser scanner configurations for robotic inspection using hyperdimensional associative memory on discrete sensing actions, improving reliability over baselines.

Reference graph

Works this paper leans on

36 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

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

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024

    arXiv 2024

  2. [22]

    Rt-1: Robotics transformer for real-world control at scale

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022

    arXiv 2022

  3. [2]

    Robomamba: Efficient vision- language-action model for robotic reasoning and manipulation

    Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Pengju An, Xiaoqi Li, Kaichen Zhou, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Efficient vision- language-action model for robotic reasoning and manipulation. Advances in Neural Information Processing Systems, 37:40085–40110, 2024

    work page 2024

  4. [3]

    Bridgedata v2: A dataset for robot learning at scale

    Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen- Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723–1736. PMLR, 2023

    2023

  5. [4]

    Open x- embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0

    Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x- embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages 6892–6903. IEEE, 2024

    work page 2024

  6. [5]

    Integrating laser profile sensor to an industrial robotic arm for improving quality inspection in manufacturing pro- cesses

    Toufik Al Khawli, Muddasar Anwar, Dongming Gan, and Shafiqul Islam. Integrating laser profile sensor to an industrial robotic arm for improving quality inspection in manufacturing pro- cesses. Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 235(1):4–17, 2021

    work page 2021

  7. [6]

    Introduction to laser ranging, profiling, and scanning

    Gordon Petrie and Charles K Toth. Introduction to laser ranging, profiling, and scanning. In Topographic laser ranging and scanning, pages 1–28. CRC Press, 2018

    work page 2018

  8. [7]

    Hand-eye calibration for 2d laser profile scanners using straight edges of common objects

    Jing Xu, Jian Li Hoo, Stylianos Dritsas, and Javier Gomez Fernandez. Hand-eye calibration for 2d laser profile scanners using straight edges of common objects. Robotics and Computer- Integrated Manufacturing, 73:102221, 2022

    work page 2022

Show all 36 references
  1. [8]

    Univla: Learning to act anywhere with task-centric latent actions, 2025

    Qingwen Bu, Yanting Yang, Jisong Cai, Shenyuan Gao, Guanghui Ren, Maoqing Yao, Ping Luo, and Hongyang Li. Univla: Learning to act anywhere with task-centric latent actions, 2025

    2025

  2. [9]

    Graspvla: a grasping foundation model pre-trained on billion-scale synthetic action data, 2025

    Shengliang Deng, Mi Yan, Songlin Wei, Haixin Ma, Yuxin Yang, Jiayi Chen, Zhiqi Zhang, Taoyu Yang, Xuheng Zhang, Heming Cui, Zhizheng Zhang, and He Wang. Graspvla: a grasping foundation model pre-trained on billion-scale synthetic action data, 2025

    2025

  3. [10]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y . Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin L...

    2025

  4. [11]

    Cot-vla: Visual chain-of-thought reasoning for vision-language- action models, 2025

    Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, Ankur Handa, Ming-Yu Liu, Donglai Xiang, Gordon Wetzstein, and Tsung-Yi Lin. Cot-vla: Visual chain-of-thought reasoning for vision-language- action models, 2025

    2025

  5. [12]

    Pointvla: Injecting the 3d world into vision-language-action models, 2025

    Chengmeng Li, Junjie Wen, Yan Peng, Yaxin Peng, Feifei Feng, and Yichen Zhu. Pointvla: Injecting the 3d world into vision-language-action models, 2025

    2025

  6. [13]

    OpenAI. GPT-4.1. https://openai.com/index/gpt-4-1/ , 2025. Accessed: 2025-05-13

    2025

  7. [14]

    Openai o3 and o4-mini system card

    OpenAI. Openai o3 and o4-mini system card. Technical report, OpenAI, 2025. Accessed: 2025-05-13

    2025

  8. [15]

    Gemini 2.5 pro preview model card

    Google. Gemini 2.5 pro preview model card. Technical report, Google, 2025. Accessed: 2025-05-13. 10

    2025

  9. [16]

    Gemini 2.5 flash preview model card

    Google. Gemini 2.5 flash preview model card. Technical report, Google, 2025. Accessed: 2025-05-13

    2025

  10. [17]

    Roboturk: A crowdsourcing platform for robotic skill learning through imitation

    Ajay Mandlekar, Yuke Zhu, Animesh Garg, Jonathan Booher, Max Spero, Albert Tung, Julian Gao, John Emmons, Anchit Gupta, Emre Orbay, et al. Roboturk: A crowdsourcing platform for robotic skill learning through imitation. In Conference on Robot Learning, pages 879–893. PMLR, 2018

    2018

  11. [18]

    Multiple interactions made easy (mime): Large scale demonstrations data for imitation

    Pratyusha Sharma, Lekha Mohan, Lerrel Pinto, and Abhinav Gupta. Multiple interactions made easy (mime): Large scale demonstrations data for imitation. In Conference on robot learning, pages 906–915. PMLR, 2018

    2018

  12. [19]

    Robonet: Large-scale multi-robot learning

    Sudeep Dasari, Frederik Ebert, Stephen Tian, Suraj Nair, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Sergey Levine, and Chelsea Finn. Robonet: Large-scale multi-robot learning. arXiv preprint arXiv:1910.11215, 2019

    1910 arXiv

  13. [20]

    Bridge data: Boosting generalization of robotic skills with cross-domain datasets

    Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021

    2021 arXiv

  14. [21]

    Bc-z: Zero-shot task generalization with robotic imitation learning

    Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Conference on Robot Learning, pages 991–1002. PMLR, 2022

    2022

  15. [23]

    Mt-opt: Continuous multi-task robotic reinforcement learning at scale

    Dmitry Kalashnikov, Jacob Varley, Yevgen Chebotar, Benjamin Swanson, Rico Jonschkowski, Chelsea Finn, Sergey Levine, and Karol Hausman. Mt-opt: Continuous multi-task robotic reinforcement learning at scale. arXiv preprint arXiv:2104.08212, 2021

    2021 arXiv

  16. [24]

    Robohive: A unified framework for robot learning

    Vikash Kumar, Rutav Shah, Gaoyue Zhou, Vincent Moens, Vittorio Caggiano, Abhishek Gupta, and Aravind Rajeswaran. Robohive: A unified framework for robot learning. Advances in Neural Information Processing Systems, 36:44323–44340, 2023

    2023

  17. [25]

    Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot

    Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023

    2023 arXiv

  18. [26]

    Droid: A large-scale in-the-wild robot manipulation dataset

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.1...

    2024 arXiv

  19. [27]

    Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation

    Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877, 2024

    2024 arXiv

  20. [28]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021

    2021

  21. [29]

    Sigmoid loss for language image pre-training, 2023

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023

    2023

  22. [30]

    Siglip 2: Multilingual vision- language encoders with improved semantic understanding, localization, and dense features, 2025

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alab- dulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. Siglip 2: Multilingual vision- language ...

    2025

  23. [31]

    Emerging properties in self-supervised vision transformers, 2021

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers, 2021

    2021

  24. [32]

    Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas...

    2023

  25. [33]

    Coverage path planning for robotic laser surfacing applications based on 3d scanning

    Ondˇrej Vaníˇcek, Michal Chaluš, and Jind ˇrich Liška. Coverage path planning for robotic laser surfacing applications based on 3d scanning. In 2024 21st International Conference on Mechatronics-Mechatronika (ME), pages 1–6. IEEE, 2024

    2024

  26. [34]

    Pso-based optimal coverage path planning for surface defect inspection of 3c components with a robotic line scanner

    Hongpeng Chen, Shengzeng Huo, Muhammad Muddassir, Hoi-Yin Lee, Yuli Liu, Junxi Li, Anqing Duan, Pai Zheng, and David Navarro-Alarcon. Pso-based optimal coverage path planning for surface defect inspection of 3c components with a robotic line scanner. IEEE Transactions on Instr...

    2025

  27. [35]

    Rein- forcement learning approach to optimizing profilometric sensor trajectories for surface inspec- tion

    Sara Roos-Hoefgeest, Mario Roos-Hoefgeest, Ignacio Álvarez, and Rafael C González. Rein- forcement learning approach to optimizing profilometric sensor trajectories for surface inspec- tion. Sensors (Basel, Switzerland), 25(7):2271, 2025

    2025

  28. [36]

    Scan the top surface of the white cube

    Keyence Corporation. LJ-X8200 High-Speed 2D/3D Laser Profiler. https://www.keyence. com/products/measure/laser-2d/lj-x8000/models/lj-x8200/ , 2025. Accessed: May 5, 2025. 12 A Technical Appendices and Supplementary Material A.1 Feature Level Ground Truth Scan Visualizations To...

    2025

Pith tools

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