Blok, Pieter M., Wang, Haozhou, Suh, Hyun Kwon, Wang, Peicheng, Burridge, James, Guo, Wei · Dec 30, 2025
arXiv cs.AI: PointRAFT uses harvester geometry to estimate potato weight from partial scans
Markdown · JSON · Original source
The authors of PointRAFT argue that incomplete scans can mislead estimates of potato weight, but that a potato harvester offers a useful geometric cue: its fixed camera-to-conveyor distance makes tuber height derivable from a point cloud. Their network adds that height as a dedicated height embedding, using it alongside the partial scan to predict weight. The authors collected 26,688 point clouds of 859 potato tubers across four cultivars and three growing seasons. On a of 5,254 point clouds representing 172 tubers, PointRAFT outperformed six point cloud regression networks without a height embedding. The authors report a mean absolute error of 11.7 g and a root mean squared error of 17.1 g. Preprocessing and inference took 6.3 ms per point cloud, which the authors say is fast enough for high-throughput estimation on a commercial potato harvester. The result is specific to the reported test set and setup; the abstract does not establish performance on other harvesters or crops. The authors also describe the network as applicable to other 3D phenotyping and robotic perception tasks, but the abstract reports potato-weight results. Code, network weights, and part of the dataset are publicly available.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Dec 30, 2025 (2025-12-30).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T02:23:02.708Z. First observed: 2026-11-05T05:03:32.545Z. Retrieved: 2026-11-10T02:09:37.809Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported performance is on a of 5,254 point clouds representing 172 tubers; the abstract does not establish performance on other harvesters. The abstract reports potato-weight results; it does not provide results for the other 3D phenotyping or robotic perception tasks the authors say the network may apply to.
Open question
How does PointRAFT’s performance change on potato harvesters beyond the one used to collect the reported data?
Source-stated
PointRAFT uses a dedicated height embedding to incorporate the height cue available from the fixed camera-to-conveyor distance on a potato harvester.
the fixed distance between the camera and the conveyor belt allows tuber height to be derived from the point cloud. To exploit this geometric cue, we introduce PointRAFT
Source-stated
On the reported test set, PointRAFT outperformed six point cloud regression networks without a height embedding, with a mean absolute error of 11.7 g and a root mean squared error of 17.1 g.
PointRAFT achieved a mean absolute error (MAE) of 11.7 g and a root mean squared error (RMSE) of 17.1 g, outperforming six point cloud regression networks without a height embedding.
Source-stated
The authors report preprocessing and inference time of 6.3 ms per point cloud and say this is fast enough for high-throughput weight estimation on a commercial potato harvester.
Requiring only 6.3 ms per point cloud for preprocessing and inference, PointRAFT is fast enough for high-throughput weight estimation on a commercial potato harvester.
Source-stated
The reported test set comprised 5,254 point clouds representing 172 tubers, within a study dataset collected across four cultivars and three growing seasons.
we collected 26,688 point clouds of 859 potato tubers on an operational harvester, spanning four cultivars and three growing seasons. On a test set of 5,254 point clouds representing 172 tubers
Source and contributor context
Blok, Pieter M., Wang, Haozhou, Suh, Hyun Kwon, Wang, Peicheng, Burridge, James, Guo, Wei · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Held-out set
Data kept aside during training and used only to test how well a model works on examples it has not seen.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 1f8011b8-81c4-4683-8be2-c6564761a6bb.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T02:23:02.708Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4e0c54f5078b0e1dd6d57adfcdd21d32ce232e68bf99fba1c8b59546adec1b4a. Source text hash: 83eebf3649a2eb3c9c57469713c058417d753c6c853ce98b7a81ebe002f57e4b.
Correction reason: The summary now separates the full collection’s cultivar and season coverage from the test set, which the abstract does not say spans all four cultivars and three seasons.
Revision and source-check history
Version 1 · 2026-11-10T02:23:02.708Z · The summary now separates the full collection’s cultivar and season coverage from the test set, which the abstract does not say spans all four cultivars and three seasons.
arXiv cs.AI: PointRAFT uses harvester geometry to estimate potato weight from partial scans
The authors of PointRAFT argue that incomplete scans can mislead estimates of potato weight, but that a potato harvester offers a useful geometric cue: its fixed camera-to-conveyor distance makes tuber height derivable from a point cloud. Their network adds that height as a dedicated height embedding, using it alongside the partial scan to predict weight. The authors collected 26,688 point clouds of 859 potato tubers across four cultivars and three growing seasons. On a of 5,254 point clouds representing 172 tubers, PointRAFT outperformed six point cloud regression networks without a height embedding. The authors report a mean absolute error of 11.7 g and a root mean squared error of 17.1 g. Preprocessing and inference took 6.3 ms per point cloud, which the authors say is fast enough for high-throughput estimation on a commercial potato harvester. The result is specific to the reported test set and setup; the abstract does not establish performance on other harvesters or crops. The authors also describe the network as applicable to other 3D phenotyping and robotic perception tasks, but the abstract reports potato-weight results. Code, network weights, and part of the dataset are publicly available.
Original source
cs.AI · Article
Source details
Original source date: Dec 30, 2025 (2025-12-30).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T02:23:02.708Z. First observed: 2026-11-05T05:03:32.545Z. Retrieved: 2026-11-10T02:09:37.809Z.
Limits
The reported performance is on a of 5,254 point clouds representing 172 tubers; the abstract does not establish performance on other harvesters. The abstract reports potato-weight results; it does not provide results for the other 3D phenotyping or robotic perception tasks the authors say the network may apply to.
Open question
How does PointRAFT’s performance change on potato harvesters beyond the one used to collect the reported data?
Source-stated
PointRAFT uses a dedicated height embedding to incorporate the height cue available from the fixed camera-to-conveyor distance on a potato harvester.
the fixed distance between the camera and the conveyor belt allows tuber height to be derived from the point cloud. To exploit this geometric cue, we introduce PointRAFT
Source-stated
On the reported test set, PointRAFT outperformed six point cloud regression networks without a height embedding, with a mean absolute error of 11.7 g and a root mean squared error of 17.1 g.
PointRAFT achieved a mean absolute error (MAE) of 11.7 g and a root mean squared error (RMSE) of 17.1 g, outperforming six point cloud regression networks without a height embedding.
Source-stated
The authors report preprocessing and inference time of 6.3 ms per point cloud and say this is fast enough for high-throughput weight estimation on a commercial potato harvester.
Requiring only 6.3 ms per point cloud for preprocessing and inference, PointRAFT is fast enough for high-throughput weight estimation on a commercial potato harvester.
Source-stated
The reported test set comprised 5,254 point clouds representing 172 tubers, within a study dataset collected across four cultivars and three growing seasons.
we collected 26,688 point clouds of 859 potato tubers on an operational harvester, spanning four cultivars and three growing seasons. On a test set of 5,254 point clouds representing 172 tubers
Source and contributor context
Blok, Pieter M., Wang, Haozhou, Suh, Hyun Kwon, Wang, Peicheng, Burridge, James, Guo, Wei · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Held-out set
Data kept aside during training and used only to test how well a model works on examples it has not seen.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 1f8011b8-81c4-4683-8be2-c6564761a6bb.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T02:23:02.708Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4e0c54f5078b0e1dd6d57adfcdd21d32ce232e68bf99fba1c8b59546adec1b4a. Source text hash: 83eebf3649a2eb3c9c57469713c058417d753c6c853ce98b7a81ebe002f57e4b.
Correction reason: The summary now separates the full collection’s cultivar and season coverage from the test set, which the abstract does not say spans all four cultivars and three seasons.
passed · 2026-11-10T02:23:02.708Z · anthropic/claude-opus-5.5
Publication: 1f8011b8-81c4-4683-8be2-c6564761a6bb
attribution — pass: The reading attributes PointRAFT, the hypothesis, the results and the broader-applicability claim to the paper's authors. It labels the venue as arXiv cs.AI, which matches the publisher field and the listed subjects.
Authors:Pieter M. Blok, Haozhou Wang, Hyun Kwon Suh, Peicheng Wang, James Burridge, Wei Guo
summary — pass: The title, summary and question track the abstract faithfully. The summary leads with the problem (incomplete scans from self-occlusion), the geometric cue (fixed camera-to-conveyor distance) and the method (height embedding), then gives results and limits. Plain wording such as 'incomplete scans' and 'partial scan' gives lay readers enough context for point cloud. The question about other harvesters follows from the single-harvester data collection and is not leading.
point clouds of crops are typically incomplete due to self-occlusion, leading to systematic errors when estimating these traits
date — pass: The originalPublished date of 2025-12-30 matches the v1 submission date. It is not the 4 Nov 2026 v2 revision or the November 2026 feed date.
[Submitted on 30 Dec 2025 (v1), last revised 4 Nov 2026 (this version, v2)]
support — pass: All figures match the abstract: 26,688/859, the 5,254/172 test set, 11.7 g MAE, 17.1 g RMSE, six baselines and 6.3 ms. The claim that this speed suffices for harvester use is kept as the authors' statement. The reading correctly separates full-dataset coverage from the test set. It flags the broader-task applicability as unshown in the abstract and does not overstate the embedding's causal contribution. The open-availability claim is supported.
The code, network weights, and a subset of the dataset are publicly available at this https URL.
media — pass: The selected media is the verified direct link to the original arXiv abstract page. It is the supplied candidate and the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Data kept aside during training and used only to test how well a model works on examples it has not seen.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivData kept aside during training and used only to test how well a model works on examples it has not seen.
Definition referenceData kept aside during training and used only to test how well a model works on examples it has not seen.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivData kept aside during training and used only to test how well a model works on examples it has not seen.
Definition reference