UNCLE-Grasp: Uncertainty-Aware Grasping of Leaf-Occluded Strawberries

TL;DR

The paper tackles robotic strawberry harvesting under severe leaf occlusion, where partial observations produce multiple plausible 3D shapes and undermine deterministic grasp planning. It introduces UNCLE-Grasp, an uncertainty-aware pipeline that samples multiple completed shapes via Monte Carlo dropout, generates candidate grasps for each completion with CGNet, and uses a physically grounded, uncertainty-aware filtering framework along with a conservative $\mathrm{LCB}$ decision rule to abstain when geometry is too uncertain. Key contributions include (i) integrating learned shape completion with MC dropout to propagate uncertainty into both local grasp filtering and object-level abstention, (ii) a multi-stage filtering scheme (global/local uncertainty, approach direction, vertical orientation, jaw clearance) and (iii) aggregation across completions using the $\epsilon$-based force-closure metric and $\mathrm{LCB}$ to guide grasp execution. Experiments in simulation and on a physical robot demonstrate that explicit handling of completion uncertainty improves robustness under occlusion by abstaining on high-risk targets while maintaining effective grasps when geometry is reliable, with practical runtime optimizations enabling real-time deployment.

Abstract

Robotic strawberry harvesting is challenging under partial occlusion, where leaves induce significant geometric uncertainty and make grasp decisions based on a single deterministic shape estimate unreliable. From a single partial observation, multiple incompatible 3D completions may be plausible, causing grasps that appear feasible on one completion to fail on another. We propose an uncertainty-aware grasping pipeline for partially occluded strawberries that explicitly models completion uncertainty arising from both occlusion and learned shape reconstruction. Our approach uses point cloud completion with Monte Carlo dropout to sample multiple shape hypotheses, generates candidate grasps for each completion, and evaluates grasp feasibility using physically grounded force-closure-based metrics. Rather than selecting a grasp based on a single estimate, we aggregate feasibility across completions and apply a conservative lower confidence bound (LCB) criterion to decide whether a grasp should be attempted or safely abstained. We evaluate the proposed method in simulation and on a physical robot across increasing levels of synthetic and real leaf occlusion. Results show that uncertainty-aware decision making enables reliable abstention from high-risk grasp attempts under severe occlusion while maintaining robust grasp execution when geometric confidence is sufficient, outperforming deterministic baselines in both simulated and physical robot experiments.

Figures (5)

• Figure 1: Overview of UNCLE-Grasp: a) Increasing leaf occlusion induces centroid shift in the reconstructed point cloud. b) Completion uncertainty from MC shape completions increases with occlusion, reducing goemetric confidence; purple regions indicate high uncertainty. c) Instead of selecting a single “best” grasp, UNCLE-Grasp evaluates object-level risk using LCB and either attempts to grasp the strawberry or abstains to avoid damaging it.
• Figure 2: Overview of the proposed uncertainty-aware strawberry grasping pipeline. Top: RGB input showing multiple strawberries partially occluded by leaves is detected and segmented, then 3D projected with the depth image to obtain a partial strawberry point cloud; A transformer-based point cloud completion network PointAttn reconstructs the missing geometry to produce the completed point cloud; MC dropout is enabled to generate K plausible completions $\{\mathcal{P}_{\text{completed}}^{(k)}\}_{k=1}^{K}$ and estimate geometric uncertainty. Middle: A grasps generation model contact-graspnet then takes each completed point cloud, $\mathcal{P}_{\text{completed}}^{(k)}$, and generates M grasps; The computed uncertainty from dropout is used to abstain from highly uncertain strawberries and filter through the grasp proposals using a) global and b) local uncertainty thresholds. Bottom: The remaining grasps are filtered using (c) orientation constraints (front-facing and non-vertical) and (d) jaw–object intersection checks. The filtered grasps from each completion sample k form the set $M'$, which are aggregated across all completion samples to compute an LCB-based feasibility score. A strawberry is attempted when LCB$>$0; otherwise, the system abstains, enabling a higher grasp success rate.
• Figure 3: (Left) Physical robot setup replicating an indoor greenhouse strawberry plantation. (Right) Simulated strawberry field in NVIDIA Isaac Sim. The strawberry plant, Unitree Z1 robotic arm, grasping mechanism, and Intel RealSense D435i RGB-D camera are annotated.
• Figure 4: Completion uncertainty under increasing occlusion. Bars indicate mean uncertainty across MC-dropout samples and whiskers denote standard deviation.
• Figure 5: Simulation-only ablations showing grasp success rate under increasing occlusion. Shape completion and geometric filtering improve robustness but remain insufficient under severe ambiguity. Values are reported as mean $\pm$ standard deviation over ten runs.