PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Grasp Targeting from Point Clouds for Log Pile Clearing on a Hydraulic Crane

George Sideris, Lucas Bessai, Heshan Fernando, Elie Ayoub, Nicolas Lemieux, Inna Sharf

arXiv:2610.07613v1cs.ROcs.LG

TL;DR

Sequential clearing of dense log piles requires repeated bundle grasps from changing scenes, while selecting grasps from unsegmented clouds must also handle rack structure. The paper trains a point-cloud grasp policy in simulation and runs it unchanged on a crane; BC and BC→RL recover more pooled inventory than the geometric heuristic, though fine-tuning’s simulated stability gain does not carry over.

  • Problem

    Dense log piles change with each bundle removal, and cloud-to-grasp learning must distinguish graspable logs from rack structure and stray returns without a separate segmentation step.

    Source: §I · p. 1

  • Method

    A shared point-cloud policy selects an observed point and predicts grasp depth and yaw, supporting behavior cloning, RL fine-tuning, and deployment.

    Source: §I · p. 1

  • Results

    BC and BC→RL recover more pooled inventory than the geometric heuristic overall, but BC→RL’s simulated stability advantage does not carry over to the crane testbed.

    Source: §VI · p. 7

  • Takeaways & Limitations

    Handling remains the sim-to-real gap, and the authors identify real-trial data or more accurate crane-dynamics modeling as next steps.

    Source: §VII · p. 8

    Evidence comes from one crane and rack, with one rebuilt pile per policy and shape and only three trials per policy.

    Source: §VI · p. 8

Abstract

from arXiv · show

In mill yards, log loaders clear dense piles by a sequence of bundle grasps: hundreds of logs rest in contact, and each removal changes the pile available to the next grasp. A learned policy chooses where to place and orient the grapple from unsegmented point clouds and runs on a trailer-mounted hydraulic forestry crane. The policy classifies at which observed point to grasp and predicts depth and grapple orientation there. The same network outputs support behavior cloning (BC), reinforcement learning (RL), and deployment. BC learns from successful top-of-pile demonstrations; RL explores for improvements by fine-tuning the cloned policy (BC$\to$RL) or by training from scratch. In simulation, BC clears 98 of 100 piles of 200 logs, while BC$\to$RL improves load stability. Twelve field trials compare a geometric heuristic, RL from scratch, BC, and BC$\to$RL through complete grasp-transport-deposit cycles. BC and BC$\to$RL deposit 93.8% and 88.9% of pooled inventory, against 80.4% for the heuristic. BC$\to$RL deposits logs on 83.6% of its cycles, against 79.6% for the heuristic and 65.7% for BC, while its simulated stability gain does not carry over to the crane testbed. Trained entirely in simulation and run unchanged on the crane, the learned policies clear more than the hand-filtered heuristic while observing unfiltered clouds that still contain the storage rack's rails and poles.

I. INTRODUCTION

The paper studies sequential bundle grasping from unsegmented point clouds, where each removal changes a dense log pile. It presents a shared learned grasp policy and compares its field clearing with a geometric heuristic.

I. INTRODUCTION

  • Dense piles change after each grasp, and learning the cloud-to-grasp mapping jointly handles material selection and grasp prediction without segmentation.Geometric filters depend on scene geometry and calibration, while a segmentation network requires its own labels.Source: §I · p. 1
  • Behavior cloning learns from successful top-of-pile demonstrations, while RL fine-tunes that policy or explores grasp sequences from scratch.Fine-tuning seeks better bundle handling; scratch training explores beyond the expert’s top-of-pile strategy.Source: §I · p. 1
  • The policy classifies which observed point to grasp and predicts depth and yaw, while a common controller executes grasp, transport, and deposit.The shared network supports behavior cloning, RL fine-tuning, and deployment.Source: §I · p. 1, Figure 2 · p. 2
  • In twelve field trials, BC and BC→RL deposit more inventory than the geometric heuristic on two of three pile shapes and overall.The clearing advantage carries over to the crane, but fine-tuning’s handling gain does not.Source: §I · p. 1

II. RELATED WORK

Prior work covers target-based log handling and grasping in smaller or looser scenes, but not sequential pile-scale clearing with hardware trials. The paper represents each grasp as a choice among observed cloud points, with associated depth and yaw.

II. RELATED WORK

  • Earlier log-grasping work targets one grasp on a static scene or small clusters; it does not address sequential pile clearing or report pile-scale hardware trials.Source: §II · p. 2
  • Unlike loose, separable clutter and pinch grasps, packed mill-yard logs are taken in bunches with a power grasp that reshapes the pile.Source: §II · p. 2

A. Grasp representation

  • At each cycle, the policy receives a point cloud, chooses grasp position and yaw, and then observes the pile changed by the fixed controller’s execution.Source: A. Grasp representation · p. 2
  • Classifying over observed points preserves separated grasp candidates that coordinate regression could average into a gap; depth offset and doubled-angle yaw shape the selected grasp.Targets stay on measured horizontal surfaces, which can include rack structure and noise, so the policy must learn graspable material.Source: A. Grasp representation · p. 2

B. Behavior cloning

Behavior cloning is trained from successful simulated top-of-pile grasps, with target depths relabeled to encourage closure on material. Its point-based targets distribute supervision across nearby observed points.

B. Behavior cloning

  • 11,366 cloud–target demonstrations from 500 simulated pile-clearing sequences supply BC training data from successful expert grasps.The privileged expert selects the highest simulated log and aligns the grapple with its axis.Source: B. Behavior cloning · p. 2
  • 0.25 m depth relabeling shifts log-center targets toward the pile surface while respecting the −1.20 m lowest-label limit.The relabeling uses a 0.056 m estimate of the surface above the log center; the resulting labels train BC and initialize BC→RL.Source: B. Behavior cloning · p. 2
  • BC assigns equal target probability to points within 0.12 m of the expert target, always including the nearest point, and combines selection, depth, yaw, and rack-pole losses.Pole columns are randomly shortened, and a penalty discourages selecting their remaining points.Source: B. Behavior cloning · pp. 2–3
  • BC training uses AdamW and retains the checkpoint with the lowest validation loss after learning-rate reductions on plateaus.Training uses a held-out 15% validation split, batch size 64, and gradient-norm clipping at 1.0.Source: B. Behavior cloning · p. 3

C. Reinforcement learning

Reinforcement learning samples grasp actions from point-cloud-conditioned outputs and optimizes rewards for bundle quantity and handling over changing pile states. Fine-tuning and scratch training use different recipes, so their comparison does not isolate BC initialization.

C. Reinforcement learning

  • RL rewards both the number of logs held and their handling, while PPO optimizes discounted returns across episodes of up to 30 cycles.Each grasp changes the pile and therefore affects later rewards; the actor sees point clouds while a privileged-state critic estimates return.Source: C. Reinforcement learning · p. 3
  • RL samples a point index, depth, and yaw encoding from the cloud, concentrating horizontal exploration on observed surfaces.The temperature sets the score distribution, while learned scales control depth and yaw exploration.Source: C. Reinforcement learning · p. 3
  • Fine-tuning freezes the BC encoder and normalization statistics and updates the head with narrow exploration, whereas scratch RL trains the full network with broader exploration.Because the recipes differ, the comparison does not isolate the effect of BC initialization.Source: C. Reinforcement learning · p. 3

IV. EXPERIMENTAL SETUP

A common engineered controller carries out the full grasp-to-deposit cycle, while the learned policy supplies the grasp decision from camera point clouds. The setup calibrates the camera and rack against crane and rack geometry.

A. Physical system and observations

  • Every target follows the same sequence of viewing, hovering, alignment, descent, closure, and lift, then transport and release above the trailer compartment.A cycle succeeds only after at least one log is deposited.Source: A. Physical system and observations · p. 3
  • 2048-point clouds are cropped to the action box plus a 0.5 m horizontal margin, retaining rails, poles, and ground in the observations.The action box covers a 2.0 × 7.0 m rack footprint and 1.4 m height; shorter clouds are zero-padded and masked.Source: A. Physical system and observations · p. 4
  • Camera calibration fits the mount transform to the grapple observed through forward kinematics, while fiducials and cloud post columns locate the rack.Simulation uses the same viewing geometry.Source: A. Physical system and observations · p. 4

B. Simulation and grasp outcomes

The study evaluates simulation and policy variants for sequential log-pile clearing, including grasp-alignment and load-stability measures. It compares the shared learned-policy heads with a geometric baseline and specifies the training protocol.

B. Simulation and grasp outcomes

  • Simulation models hundreds of contacting logs across randomized pile profiles, with camera rendering matched to the deployed sensor.The crane model includes actuated arm joints, grapple yaw and tongs, and passive suspension.Source: B. Simulation and grasp outcomes · p. 4
  • Alignment compares grapple and held-log axes, while stability scores load tilt over a one-second post-lift window.Both scores lie in [0, 1], with exponents emphasizing nearly aligned logs and nearly level loads.Source: B. Simulation and grasp outcomes · p. 4, Figure 6 · p. 4

C. Compared policies and training

  • The geometric heuristic selects a high point in a cropped cloud, uses pole exclusions and local-axis yaw, and grasps below it.Its depth is 0.25 m in simulation and field-tuned to 0.30 m on the crane.Source: C. Compared policies and training · p. 4
  • Figure 5 pairs mast-camera views from the crane and simulation using calibrated camera and rack placement.Figure 6 presents alignment and stability, with insets showing the shaped scores.Source: Figure 5 · p. 4, Figure 6 · p. 4
  • Regression BC predicts one box-normalized position and yaw, while RL from scratch can replace categorical point selection with Gaussian grasp-coordinate exploration.The regression comparison keeps demonstrations, preprocessing, backbone, and optimizer unchanged; the RL comparison shares environment and reward.Source: C. Compared policies and training · p. 4
  • Fine-tuning and RL from scratch use three seeds and 50k cycles per run, with checkpoints screened before final evaluation and deployment.The selected checkpoint is the one that empties the most racks over 40 deterministic episodes.Source: C. Compared policies and training · p. 4

D. Evaluation protocols

The evaluation combines shared-pile simulation with twelve crane trials across three pile shapes. Scoring BC completes more simulated racks than its expert, while fine-tuning improves simulated stability but trades off cycle success.

D. Evaluation protocols

  • Simulation evaluates six policies over 100 shared training-distribution pile seeds, ending at an empty rack or 30-cycle horizon.The c95 mean includes only episodes reaching 95% removal; cycles to termination count grasp decisions until the episode ends.Source: D. Evaluation protocols · p. 4
  • Crane trials test the geometric heuristic, BC, BC→RL, and RL from scratch once each on single-mound, double-mound, and flat racks.Rebuilt piles share shape categories but are not identical paired states; initial inventories range from 199 to 219 logs.Source: D. Evaluation protocols · pp. 4–5
  • Field Clear measures the share of initial inventory deposited in the trailer, while pooled campaign rates sum trial counts before calculating ratios.Knocked-off logs are not cleared, and alignment is rated from drooping or falling logs through compact, well-aligned bundles.Source: D. Evaluation protocols · p. 5

A. Imitation, exploration, and simulated clearing

  • Scoring BC finishes more simulated racks than its privileged expert, and fine-tuning reaches higher mean return and handling than learning from scratch.Unfinished expert episodes retain a median of two logs, typically isolated against the floor.Source: A. Imitation, exploration, and… · p. 5
  • Fine-tuning raises simulated stability by 0.035±0.003, approximately halves knocked-off logs, and lowers cycle success by 2.8 ± 0.8 percentage points.The reward permits a trade-off between grasping more often and carrying a more stable load.Source: A. Imitation, exploration, and… · p. 5
  • RL from scratch uses a regional strategy with median 0.4 m spacing between early targets, versus 1.5 m for the expert and fine-tuned policy, but does not improve completion.Under the tested reward and budget, it has the weakest handling.Source: A. Imitation, exploration, and… · p. 5

B. Representation and design ablations

Representation and design ablations compare point scoring with coordinate regression, alternative RL action heads, reward terms, and fine-tuning choices. Offline replay also examines how crop size affects grasp selection on crane clouds.

B. Representation and design ablations

  • The scoring head empties 98, 98, and 99 of 100 racks across three BC seeds, while the regression head empties 33, 0, and 0.The regression run in Table I retains a median 45 logs in unfinished episodes.Source: B. Representation and design ablations · p. 5
  • Across regression runs, 14.6–24.7% of decisions target locations with no observed point within 0.5 m horizontally; scoring BC has neither diagnostic failure.Regression reissues such a target within 0.15 m of the previous one in 5.1–8.6% of decisions.Source: B. Representation and design ablations · p. 5
  • The categorical RL head’s screened completion rates are 97.5%, 57.5%, and 0% across seeds, while the Gaussian head empties no rack on any seed.This comparison holds environment and reward fixed while changing categorical point selection to Gaussian coordinate exploration.Source: B. Representation and design ablations · p. 5
  • The deployed reward gives the highest mean training clearing; removing quality terms increases bundle size and lowers stability.The ablation compares the deployed reward with normalized multiplication, normalized addition, and log count alternatives.Source: B. Representation and design ablations · p. 5
  • With the encoder frozen, a critic given privileged log poses finishes with higher mean training return and a narrower late seed band than a point-cloud critic.Leaving the encoder trainable lowers training return and produces screened completion rates from 72.5% to 100%.Source: B. Representation and design ablations · p. 5
  • Offline replay shows regression predicting intermediate targets between mounds, margin-trained scoring selecting a mound point, and tight-crop scoring repeatedly selecting unseen rails and poles.The margin-trained policy was therefore selected for deployment.Source: B. Representation and design ablations · p. 5

C. Clearing and handling on the crane testbed

On the crane, BC and BC→RL deposited more inventory than the geometric heuristic overall, but no learned policy led on every measure. Fine-tuning’s simulated stability gain did not transfer to the testbed.

C. Clearing and handling on the crane testbed

D. Where physical clearing fails

Physical clearing failures include repeated targeting of rack structure, chokes on steep piles, and depth errors. The learned policies sometimes recover from structure targets, but their behavior and failure modes differ across trials.

D. Where physical clearing fails

  • All six heuristic structure/noise empty cycles occur in the final third; repeated selection of high pole returns can trap the height rule, while BC returns to material after its three flat-rack structure cycles.BC has 13 structure/noise cycles in 70, compared with six in 49 for the heuristic; BC→RL has three in 55, all on base rails rather than poles.Source: D. Where physical clearing fails · p. 7
  • BC’s double-mound trial has four chokes despite no gap-targeting cycles; on steep flanks, the successful depth band can enclose too much material for closure.Simulation shows no chokes.Source: D. Where physical clearing fails · p. 7
  • BC→RL ends its single-mound trial after thirteen successful cycles when it closes above remaining logs, but the trial does not distinguish depth prediction from residual camera or kinematic error.Source: D. Where physical clearing fails · p. 7
  • RL from scratch targets a median 0.17 m above the observed mound surface without offsets; its reported 98.0% Clear includes later depth-offset adaptation.BC and BC→RL learn depth from demonstration labels informed by crane commissioning, whereas RL from scratch learns depth only in simulation.Source: D. Where physical clearing fails · p. 7

VI. DISCUSSION

Across the campaign, BC and BC→RL recover more inventory than the geometric heuristic, but their physical strengths differ. The discussion identifies limited field evidence and a sim-to-real handling gap; the conclusion points to real-trial data or improved crane dynamics modeling as next steps.

VI. DISCUSSION

  • BC and BC→RL recover more inventory than the geometric heuristic on two pile shapes and across the pooled campaign, despite using a wider cloud without geometric support filtering or separate segmentation.Source: §VI · p. 7
  • BC recovers the most pooled inventory, while BC→RL has the highest pooled cycle success but loses its simulated stability advantage on the crane.Source: §VI · p. 7
  • The authors recommend evaluating task completion and handling separately over full pile depletion because better simulated handling need not produce better physical handling.Source: §VI · p. 8
  • Evidence comes from one crane and rack and one rebuilt pile per policy and shape; three trials per policy cannot establish cross-platform performance or explain the stability result.Source: §VI · p. 8
  • Simulation misses physical chokes and the field stability ordering; proposed next steps include measuring contact-parameter effects, using small amounts of real-trial data for depth faults, extending the rail penalty, and randomizing contact properties.Source: §VI · p. 8

VII. CONCLUSION

  • One point-selection, depth, and yaw network supports BC, RL, and deployment with unchanged weights; closing the remaining handling gap is the stated next step.The conclusion identifies real-trial data or more accurate crane-dynamics modeling as possible approaches.Source: §VII · p. 8

Related papers

Loading 2610.07613v1…