Pith. sign in

REVIEW 3 major objections 6 minor 23 references

A hierarchical reinforcement learning policy that separates obstacles before grasping can harvest occluded strawberries with 71.7–88.3% success after sim-to-real transfer.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 03:38 UTC pith:42NIBETU

load-bearing objection A useful HRL push-before-grasp system for strawberry harvesting, but the real-world statistical evidence for 'consistently outperforms' is weaker than the text suggests. the 3 major comments →

arxiv 2607.13799 v1 pith:42NIBETU submitted 2026-07-15 cs.RO

Vision-Based Obstacle Separation for Strawberry Harvesting in Clusters Using Hierarchical Reinforcement Learning

Teng Li , Hanfei Shi , Chunjiang Zhao , Ya Xiong This is my paper
classification cs.RO
keywords strawberry harvestinghierarchical reinforcement learningobstacle separationpush-and-graspvision-guided policyadaptive explorationsim-to-real transferselective harvesting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a robotic strawberry harvester can reliably pick ripe fruit even when it is tightly surrounded by unripe fruit, by explicitly splitting the job into two learned stages: first push aside the obstructing fruit, then grasp the target. The proposed hierarchical framework, VGPA, uses a vision-guided high-level policy to decide when to push versus grasp, and a low-level continuous controller with a progressive exploration strategy to execute those actions. In simulation the policy reaches 96.7% high-level success, and on a real parallel robot it outperforms direct picking under slight, moderate, and dense occlusion while adding only about 1.2 seconds per fruit. If the claim holds, it means active obstacle separation is a viable alternative to passive avoidance for clustered crops.

Core claim

The paper's central claim is that decomposing clustered strawberry harvesting into sequential obstacle separation and target grasping, and coordinating those stages with a hierarchical reinforcement learning loop, makes selective harvesting feasible under heavy occlusion. VGPA's high level selects an option (push or grasp) with the help of a vision-guided module that estimates occlusion and provides curriculum-style supervision early in training; its low level learns continuous pushing and grasping via DDPG with a Progressive Adaptive Exploration Strategy that adaptively mixes random, guided, and policy-driven actions based on stage success rates. The authors report 96.7% high-level success

What carries the argument

The central object is VGPA, a hierarchical goal-conditioned reinforcement learning framework. At the high level, a vision-guided module uses visual detection of the target and surrounding fruit to estimate occlusion and supervise option selection (push vs. grasp), with the fraction of vision-guided episodes decaying as the policy's own success rate rises. At the low level, the Progressive Adaptive Exploration Strategy (PAES) regulates exploration by mixing uniform random actions, guided actions from a previously mastered stage, and the current policy's actions, with probabilities and noise scales updated by smoothed stage success. Hindsight-experience relabeling is used to improve sample eff

Load-bearing premise

The real-world evaluation uses only 20 hand-constructed scenes per occlusion level with three repeats each (60 trials per condition) and no statistical significance testing, so the claimed consistent outperformance over direct picking assumes those scenes faithfully represent the range of cluster geometries in practice.

What would settle it

A larger field study with, say, 100 or more naturally occurring clusters per occlusion condition, reporting per-trial success rates and confidence intervals; if the success advantage over direct picking shrinks to within noise under moderate or dense occlusion, the central claim would be weakened.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Robotic harvesters can succeed on fruits that are visually inaccessible by actively clearing obstacles first, rather than only avoiding them.
  • A vision-guided high-level policy converges faster and more stably than a hierarchy without such guidance, as shown by the ablation comparisons.
  • Policies trained in simulation can transfer to a real robot without fine-tuning when trained in end-effector pose space, simplifying deployment.
  • The method offers a better success-versus-time trade-off than a vision-language-model planner: 80.0% average success in 4.15 s versus 59.7% in 2.93 s for direct picking, and 7.23 s for the VLM baseline.
  • In dense occlusion the approach still fails in about 28% of trials, often because the swallowing motion picks both the obstacle and the target fruit together.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same push-then-grasp decomposition could transfer to other cluster-growing crops such as tomatoes, grapes, or cherries, where the obstacle is also unripe fruit or foliage and the reward structure is analogous.
  • The vision-guided module acts as a learned curriculum: using a separate perception signal to supervise option selection early in training is a general recipe for long-horizon manipulation tasks where a clear visual gating signal exists.
  • The residual failures under dense occlusion suggest that a more dexterous end-effector (e.g., suction or compliant fingers) combined with the learned pushing policy could push success rates higher than the reported 71.7%.
  • Because the real-world results came from 20 researcher-built scenes per condition in a single greenhouse, the natural next test is a diverse field trial with naturally varying cluster geometries and statistical confidence intervals.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes VGPA, a hierarchical reinforcement learning framework for selective strawberry harvesting in clusters. The high-level policy selects between obstacle separation and grasping options; a vision-guided (VG) module injects image-based occlusion information to accelerate high-level learning, while a Progressive Adaptive Exploration Strategy (PAES) adaptively balances random, guided, and policy-driven low-level actions. The low-level policies are trained with DDPG and HER. The authors report a 96.7% high-level success rate in PyBullet simulation and, after sim-to-real transfer on a custom parallel robot, real-world success rates of 88.3%, 80.0%, and 71.7% under slight, moderate, and dense occlusion, compared with 75.7%, 58.3%, and 45.0% for direct picking. The paper also includes ablations of VG and PAES and a comparison against HAC and a VLM-based planner.

Significance. If the reported results hold, the paper makes a practically relevant contribution to agricultural robotics by showing that a push-before-grasp hierarchical policy can improve harvesting in clustered strawberry environments and transfer from simulation to a real robot. The hierarchical decomposition, vision-guided option supervision, and adaptive exploration are sensible design choices, and the inclusion of real-world deployment on a self-built robot is a strength. The paper also provides a clear qualitative failure analysis and an honest acknowledgment of environmental simplifications. However, the central real-world superiority claim rests on a small, clustered sample with no statistical reporting, and the HAC baseline comparison is not explained well enough to establish that the proposed method is genuinely better rather than better tuned or differently inductive-biased. The work is therefore promising but not yet fully convincing as stated.

major comments (3)
  1. [IV-B3, Table I] The claim that VGPA 'consistently outperforms direct picking under different occlusion conditions' is not statistically established. Each reported rate is based on 60 trials (20 scenes × 3 repeated trials), the three trials per scene are not independent, and no confidence intervals, per-scene breakdowns, or statistical tests are provided. For the slight-occlusion condition, 88.3% (53/60) vs 75.7% (45/60) is only marginally significant even under a naive two-proportion test (z≈1.8, p≈0.07), and clustering further reduces the effective sample size. The moderate and dense occlusion differences are larger, but the blanket claim covers all conditions. The authors should report cluster-adjusted confidence intervals and a proper hypothesis test, and specify whether the 20 scenes were randomly sampled or fixed.
  2. [IV-B1, Fig. 5] The HAC comparison reports a striking gap (96.7% vs 7.77% high-level success) without explaining why HAC essentially fails to learn. The manuscript does not describe HAC's hyperparameters, network architecture, exploration schedule, reward structure, or whether the same epoch/episode budget and random seeds were used. Because VGPA uses vision-guided demonstrations and a stage-aware exploration schedule, the comparison may not be apples-to-apples. The authors should provide a fair HAC configuration, report multiple-seed means with variance, and ideally run HAC with the same number of environment interactions and a comparable exploration mechanism to isolate the contribution of the proposed components.
  3. [III-B, Algorithm 1] The vision-guided module directly supplies the option in 'VG episodes,' with the fraction of such episodes governed by Eq. (3). This means part of the high-level training signal comes from a hand-designed occlusion rule, not from the learned policy. The reported 96.7% high-level success rate may therefore conflate the learned decision policy with the injected heuristic, especially if VG episodes or VG-informed option selection remain active during evaluation. Please state explicitly whether VG supervision is disabled at evaluation and report high-level performance with and without the vision-guided option injection.
minor comments (6)
  1. [IV-B3, Table I vs text] The text states direct picking's average grasping time is 2.93 s, but Table I reports 2.63 s. Please make the numbers consistent.
  2. [III-D, Eqs. (8)-(10)] The reward thresholds in Eqs. (8)-(10) are presented without motivation. For example, the pushing reward requires (obs-to-tar)_x to lie in a 5 mm band; please define the coordinate frame and explain why this interval was chosen. Also, Eq. (7) uses the symbol g⟨ which is never defined.
  3. [III-B, Eq. (3)] Eq. (3) contains several schedule constants (α_max, α_min, k, τ0, λ) but no sensitivity analysis is given. Since PAES and VG are central contributions, the sensitivity of the final result to these constants should be discussed or at least bounded.
  4. [III, Algorithm 1] The high-level policy is said to be optimized using 'DIOL,' but DIOL is neither expanded nor cited. Please define it in the text or replace with the full algorithm name and reference.
  5. [IV-B2, Fig. 6] The ablation curves in Fig. 6 show only one trajectory per setting; reporting multiple seeds with mean and standard deviation would substantially strengthen the claim that VG and PAES improve convergence and stability.
  6. [IV-B3, VLM comparison] The VLM baseline is described as 'developed in our group' but no implementation details or citation are provided beyond [23]. Since this is an internal comparison, it should be positioned as such and the differences in action space, perception, and planning should be described.

Circularity Check

0 steps flagged

No significant circularity: reported success rates are measured experimental outcomes, not quantities derived from fitted parameters or forced by self-citation.

full rationale

The paper's central claims are empirical: VGPA achieves 96.7% high-level success in simulation and 88.3%/80.0%/71.7% real-world success under increasing occlusion, compared with 75.7%/58.3%/45.0% for direct picking. These are measured by executing the trained policy in test scenarios, not by algebraically deriving the result from the method's parameters. The vision-guided module (Eq. 3, Algorithm 1) is a training-time curriculum that adjusts the proportion of hand-specified option demonstrations; it is not a post-hoc fit to the reported success rates, and the paper explicitly reduces its influence as the high-level success rate rises. PAES (Eqs. 4–6) adapts exploration during training; it does not produce the final success rates by construction. The self-citations ([5], [19], [22], [23]) describe a prior obstacle-separation strategy, a detection/ripeness method, the robotic arm platform, and a VLM baseline; none is invoked as a uniqueness theorem or as the justification that the proposed method's success rates must have particular values. The small real-world sample size and absence of statistical testing are correctness/evidence concerns, not circularity. No step in the derivation chain reduces to its own inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim rests on hand-chosen hyperparameters for the two novel modules (VG schedule and PAES exploration), a reward-shaping band that appears tuned to simulation, an unquantified perception assumption, and a simplified simulation model whose limits the authors concede. No new physical entities are posited. The evaluation assumptions (baseline quality, scene representativeness) also carry weight.

free parameters (3)
  • Vision-guided demonstration ratio schedule (α_max, α_min, k, τ_0, λ) = not fully reported
    Eq. (3) defines the number of guided episodes per epoch using five hand-chosen schedule constants; the paper does not state their values or a sensitivity analysis, so the reported convergence benefit could be tuned to the task.
  • PAES bounds (c_α, c_σ, c_g, γ_min, γ_max, κ) = not fully reported
    Eqs. (4)-(6) define the PAES exploration scheduler with six constants that control exploration; without values or ablations over them, the claimed stability improvement is conditional on hand-set hyperparameters.
  • Low-level reward weights and thresholds (0.7/0.3, distance thresholds 0.045/0.01, pose ±10°) = 0.7/0.3 weight, 0.04–0.045 push distance, 0.0–0.01 grasp distance, ±10°
    The low-level reward is a hand-weighted combination of sparse distance and pose rewards. The specific narrow band in Eq. (8) (0.04 ≤ x ≤ 0.045) looks like a threshold tuned in simulation; no rationale or sensitivity analysis is given.
axioms (4)
  • domain assumption YOLOv11 detection and depth back-projection produce sufficiently accurate 3D fruit positions for policy input.
    Sections III-A and IV-B3 state that severe occlusion affects perception and localization and that this caused failures, so the central success rates are conditional on detection accuracy that is not separately quantified.
  • domain assumption The PyBullet simulation with SolidWorks strawberry models captures enough physics (rigid push interactions, similar friction) for the learned push behavior to transfer.
    Section V concedes the simulation 'simplified certain characteristics of real agricultural environments, such as fruit deformation, peduncle flexibility, and complex interactions among neighboring fruits', which is exactly where the sim-to-real gap could break the claimed transfer.
  • domain assumption The benchmark HAC baseline was implemented with comparable hyperparameter and training effort to VGPA.
    Section IV-B1 reports HAC high-level success of 7.77% vs 96.7% for VGPA but provides no hyperparameter details, budget, or tuning for HAC; the near-zero performance is consistent with a weak or untuned baseline.
  • domain assumption The 20-scene occlusion taxonomy (slight/moderate/dense) is representative of real clustered strawberry canopies.
    Real-world results are averaged within the authors' own three scene classes; the same-group VLM paper [23] presumably uses a similar taxonomy, so generalization to arbitrary commercial fields is untested.
invented entities (1)
  • No new physical entities (forces, objects, or mediators) are introduced. no independent evidence
    purpose: n/a
    The paper introduces only software components (VG module, PAES); nothing new is postulated about the physical world, so the gravitons problem does not apply.

pith-pipeline@v1.3.0-alltime-deepseek · 8969 in / 8252 out tokens · 62602 ms · 2026-08-02T03:38:27.906002+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Vision-Based Obstacle Separation for Strawberry Harvesting in Clusters Using Hierarchical Reinforcement Learning." pith.science (2026). https://pith.science/paper/42NIBETU

@misc{pith2026260713799,
  author       = {Pith},
  title        = {Pith review of: Vision-Based Obstacle Separation for Strawberry Harvesting in Clusters Using Hierarchical Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42NIBETU}},
  note         = {Machine review of arXiv:2607.13799}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Selective harvesting in clustered strawberry environments is challenging because ripe fruits are often occluded by surrounding unripe fruits, making direct grasping unreliable. To address this problem, this paper proposes a hierarchical reinforcement learning framework, termed VGPA, which integrates a vision-guided decision mechanism and a Progressive Adaptive Exploration Strategy (PAES) for vision-based obstacle separation and harvesting. The task was decomposed into two sequential stages: obstacle separation and target grasping. At the high level, the vision-guided mechanism improved option selection and accelerated policy convergence. At the low level, PAES improved exploration efficiency and training stability during continuous control learning. In simulation experiments, the learned policy achieved a success rate of 96.7%. In addition, sim-to-real transfer experiments on a self-developed parallel robot showed that the proposed method achieved success rates ranging from 71.7% to 88.3%, outperforming direct picking while requiring only 1.22~s more average harvesting time. These results verified the effectiveness, generalization ability, and practical potential of the proposed method for robotic harvesting in complex clustered environments.

Figures

Figures reproduced from arXiv: 2607.13799 by Chunjiang Zhao, Hanfei Shi, Teng Li, Ya Xiong.

Figure 1
Figure 1. Figure 1: Top: A target strawberry cluster in the initial observation, where [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the proposed VGPA framework. The RGB-D observation was first processed by YOLOv11 to detect and localize the target [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The structure of VG at the high level. C. Progressive Adaptive Exploration Strategy (PAES) An adaptive exploration strategy, PAES, was designed for hierarchical reinforcement learning to regulate exploration behaviors according to the learning progress of different manipulation stages. As discussed above, when training different levels of policies in parallel within a hierarchical reinforcement learning sy… view at source ↗
Figure 6
Figure 6. Figure 6: The ablation results for VG and PAES modules. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Field test of the strawberry harvesting robot (HarvestFlex) in a [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Success rate training curves for HAC and VGPA. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Real-world harvesting examples under different occlusion conditions. Rows (a), (b), and (c) correspond to slight, moderate, and dense occlusion [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 4 linked inside Pith

  1. [1]

    Machine learning in agriculture domain: A state-of-art survey.Artificial Intelligence in the Life Sciences, 1:100010, 2021

    Vishal Meshram, Kailas Patil, Vidula Meshram, Dinesh Hanchate, and SD Ramkteke. Machine learning in agriculture domain: A state-of-art survey.Artificial Intelligence in the Life Sciences, 1:100010, 2021

    2021

  2. [2]

    Fruit localization and environment perception for strawberry harvest- ing robots.IEEE access, 7:147642–147652, 2019

    Yuanyue Ge, Ya Xiong, Gabriel Lins Tenorio, and P ˚al Johan From. Fruit localization and environment perception for strawberry harvest- ing robots.IEEE access, 7:147642–147652, 2019

    2019

  3. [3]

    Peduncle collision-free grasping based on deep reinforcement learning for tomato harvesting robot.Computers and Electronics in Agriculture, 216:108488, 2024

    Yajun Li, Qingchun Feng, Yifan Zhang, Chuanlang Peng, Yuhang Ma, Cheng Liu, Mengfei Ru, Jiahui Sun, and Chunjiang Zhao. Peduncle collision-free grasping based on deep reinforcement learning for tomato harvesting robot.Computers and Electronics in Agriculture, 216:108488, 2024

    2024

  4. [4]

    Reaching pruning locations in a vine using a deep rein- forcement learning policy

    Francisco Yandun, Tanvir Parhar, Abhisesh Silwal, David Clifford, Zhiqiang Yuan, Gabriella Levine, Sergey Yaroshenko, and George Kantor. Reaching pruning locations in a vine using a deep rein- forcement learning policy. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 2400–2406. IEEE, 2021

    2021

  5. [5]

    An improved obsta- cle separation method using deep learning for object detection and tracking in a hybrid visual control loop for fruit picking in clusters

    Ya Xiong, Yuanyue Ge, and P ˚al Johan From. An improved obsta- cle separation method using deep learning for object detection and tracking in a hybrid visual control loop for fruit picking in clusters. Computers and Electronics in Agriculture, 191:106508, 2021

    2021

  6. [6]

    Zero-shot sim-to- real reinforcement learning for fruit harvesting.arXiv preprint arXiv:2505.08458, 2025

    Emlyn Williams and Athanasios Polydoros. Zero-shot sim-to- real reinforcement learning for fruit harvesting.arXiv preprint arXiv:2505.08458, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Efficient learning of goal-oriented push-grasping synergy in clutter

    Kechun Xu, Hongxiang Yu, Qianen Lai, Yue Wang, and Rong Xiong. Efficient learning of goal-oriented push-grasping synergy in clutter. IEEE Robotics and Automation Letters, 6(4):6337–6344, 2021

    2021

  8. [8]

    Learning bifunctional push-grasping synergistic strategy for goal- agnostic and goal-oriented tasks

    Dafa Ren, Shuang Wu, Xiaofan Wang, Yan Peng, and Xiaoqiang Ren. Learning bifunctional push-grasping synergistic strategy for goal- agnostic and goal-oriented tasks. In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2909–

  9. [9]

    Learning extrinsic dexterity with parameterized manipu- lation primitives

    Shih-Min Yang, Martin Magnusson, Johannes A Stork, and Todor Stoyanov. Learning extrinsic dexterity with parameterized manipu- lation primitives. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5404–5410. IEEE, 2024

    2024

  10. [10]

    Between mdps and semi-mdps: A framework for temporal abstraction in reinforce- ment learning.Artificial intelligence, 112(1-2):181–211, 1999

    Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforce- ment learning.Artificial intelligence, 112(1-2):181–211, 1999

    1999

  11. [11]

    The option-critic architecture

    Pierre-Luc Bacon, Jean Harb, and Doina Precup. The option-critic architecture. InProceedings of the AAAI conference on artificial intelligence, volume 31, 2017

    2017

  12. [12]

    Hierarchical reinforce- ment learning by discovering intrinsic options.arXiv preprint arXiv:2101.06521, 2021

    Jesse Zhang, Haonan Yu, and Wei Xu. Hierarchical reinforce- ment learning by discovering intrinsic options.arXiv preprint arXiv:2101.06521, 2021

    Pith/arXiv arXiv 2021

  13. [13]

    Hierarchical reinforcement learning with universal policies for multistep robotic manipulation.IEEE Transac- tions on Neural Networks and Learning Systems, 33(9):4727–4741, 2021

    Xintong Yang, Ze Ji, Jing Wu, Yu-Kun Lai, Changyun Wei, Guoliang Liu, and Rossitza Setchi. Hierarchical reinforcement learning with universal policies for multistep robotic manipulation.IEEE Transac- tions on Neural Networks and Learning Systems, 33(9):4727–4741, 2021

    2021

  14. [14]

    Hierarchical visual policy learning for long-horizon robot manipulation in densely cluttered scenes

    Hecheng Wang, Lizhe Qi, Ziheng Wang, Jiankun Ren, Wei Li, and Yunquan Sun. Hierarchical visual policy learning for long-horizon robot manipulation in densely cluttered scenes. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 1149–1155. IEEE, 2025

    2025

  15. [15]

    Impedance primitive-augmented hierarchical reinforcement learning for sequential tasks

    Amin Berjaoui Tahmaz, Ravi Prakash, and Jens Kober. Impedance primitive-augmented hierarchical reinforcement learning for sequential tasks. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 10973–10979. IEEE, 2025

    2025

  16. [16]

    Context- aware grasp generation in cluttered scenes

    Dinh-Cuong Hoang, Johannes A Stork, and Todor Stoyanov. Context- aware grasp generation in cluttered scenes. In2022 International Conference on Robotics and Automation (ICRA), pages 1492–1498. IEEE, 2022

    2022

  17. [17]

    Learning to grasp the ungraspable with emergent extrinsic dexterity

    Wenxuan Zhou and David Held. Learning to grasp the ungraspable with emergent extrinsic dexterity. InConference on Robot Learning, pages 150–160. PMLR, 2023

    2023

  18. [18]

    Continuous control with deep reinforcement learning.arXiv preprint arXiv:1509.02971, 2015

    Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning.arXiv preprint arXiv:1509.02971, 2015

    Pith/arXiv arXiv 2015

  19. [19]

    Light- resilient visual regression of strawberry ripeness for robotic harvesting

    Meili Sun, Shimin Hu, Chunjiang Zhao, and Ya Xiong. Light- resilient visual regression of strawberry ripeness for robotic harvesting. Computers and Electronics in Agriculture, 241:111169, 2026

    2026

  20. [20]

    Pybullet quickstart guide.ed: PyBullet Quickstart Guide

    Erwin Coumans and Yunfei Bai. Pybullet quickstart guide.ed: PyBullet Quickstart Guide. https://docs. google. com/document/u/1/d, 2021

    2021

  21. [21]

    Learning multi-level hierarchies with hindsight.arXiv preprint arXiv:1712.00948, 2017

    Andrew Levy, George Konidaris, Robert Platt, and Kate Saenko. Learning multi-level hierarchies with hindsight.arXiv preprint arXiv:1712.00948, 2017

    Pith/arXiv arXiv 2017

  22. [22]

    Design and control of a novel six-degree-of-freedom hybrid robotic arm

    Yang Chen, Zhonghua Miao, Yuanyue Ge, Sen Lin, Liping Chen, and Ya Xiong. Design and control of a novel six-degree-of-freedom hybrid robotic arm. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3597–3604. IEEE, 2024

    2024

  23. [23]

    Active obstacle separation: Vision-language model (vlm) driven clearing decisions for robotic harvesting.Smart Agricultural Technology, page 102037, 2026

    Quan Zhao, Yuanyue Ge, Hui Zhao, and Ya Xiong. Active obstacle separation: Vision-language model (vlm) driven clearing decisions for robotic harvesting.Smart Agricultural Technology, page 102037, 2026

    2026