Sim-to-Real Reinforcement Learning for Path-Safe Target Reaching in Robotic Harvesting of Hydroponic Lettuce
Authors: Al Bashir, Partho Ghose, Yibin Wang, Robert G. Hardin, Joshua Peeples, Genhua Niu, Azlan Zahid
Organizations: Department of Biological & Agricultural Engineering, Texas A&M University, College Station, TX, 77843, USA · Texas A&M AgriLife Research, Texas A&M University System, Dallas, TX, 75252, USA · Department of Electrical & Computer Engineering, Texas A&M University, College Station, TX, 77843, USA · Department of Horticultural Sciences, Texas A&M University, College Station, TX, 77843, USA
Robotic harvesting requires precise manipulation to ensure accurate positioning of the end-effector, especially in a constraint workspace. This study introduces the Path-Safe Approach Proximal Policy Optimization (PSA-PPO), a target-conditioned local reaching framework designed for robotic harvesting of nutrient film technique (NFT)-hydroponic lettuce in greenhouses. The proposed PSA-PPO preserves the standard clipped PPO objective while introducing a harvesting-oriented policy formulation that integrates target-conditioned observations, corridor-aware spatial constraints, trajectory-history-based safety encoding, and task-specific success criteria for constrained robotic manipulation. PSAPPO was trained entirely in Isaac Lab and successfully transferred to a physical UR10e robotic arm without any real-world policy fine-tuning. In a controlled physical benchmark against 150 scheduled targets, PSA-PPO achieved the highest reaching rate of 94.7%, outperforming Cartesian+TOTG (86.0%), RRTConnect (88.0%), and RRT* (80.0%), while reducing median RMS jerk by 39.8 to 73.0% relative to the classical planners during the final approach segment. A formulation ablation evaluated over 200 simulation targets demonstrates that path-safety and side-corridor shaping are the principal drivers of harvesting-compatible performance: removing either term substantially degrades strict success even when terminal endpoint accuracy is preserved or improved, confirming that optimization toward the endpoint alone is an insufficient objective for constrained crop-proximal manipulation. Finally, PSA-PPO was integrated with an RGB-D vision module and an end-effector and evaluated as a complete greenhouse harvesting system across 144 lettuce harvesting attempts, achieving 94.4% vision-guided target reaching and 86.1% complete harvesting success.
Figures & tables
Figure 1: Overview of the proposed sim-to-real PSA-PPO harvesting framework. The robot-mounted RGB-D camera provides RGB imagery and aligned depth to the lettuce perception module, which generates temporally smoothed 3D cutting-point targets. PSA-PPO is trained entirely in Isaac Lab for target-conditioned, path-safe side-approach reaching. During physical deployment, the frozen policy and vision-derived targets are integrated through the ROS action-control layer to command the UR10e for side-approach harvesting of nutrient film technique (NFT)-grown lettuce.
Figure 2: Training architecture of the path-safe, side-corridor approach policy (PSA-PPO) for the UR10e manipulator. Each target pT defines a horizontal approach direction aT and same-height staging reference pS and determines a geometry-conditioned safe-start configuration. The actor receives the 42-dimensional policy observation ξt and produces a six-dimensional action uˉt , which is bounded, filtered, and scaled before application. The environment evaluates target closure, corridor following, tool orientation, wrist–TCP path safety, and motion regularity. Rollout data and critic estimates are used to compute returns and advantages for the clipped PPO update. The selected actor checkpoint is subsequently validated in simulation and exported for physical deployment.
Figure 3: Redesigned cut-and-hold lettuce harvesting end effector. (a–f) SolidWorks components and assembled geometry, including top, bottom, side, open, and closed configurations; (g) fabricated mechanism mounted on the OnRobot 2FG7 and UR10e manipulator.
Figure 4: ROS-based harvesting-control architecture. Smoothed RGB-D cutting points are transformed to the robot frame and corrected for end-effector geometry. The harvest controller performs target-conditioned pre-positioning before activating the PSA-PPO policy for the learned local side approach. Policy outputs are mapped to UR10e joint commands through the ROS trajectory interface, with independent hardware-side timing and acceleration safeguards. A separate gripper-control branch actuates the 2FG7-driven cut-and-hold mechanism and releases harvested lettuce at the collection basket.
Evaluation
Population
Primary outcomes
Simulation validation
200 workspace targets
Reaching the target w/o any safety measures, path-safe reaching, strict success, final TCP error, and maximum below-plane violation
Formulation ablation
Same 200 targets per formulation
Reach without safety measures, path-safe, and strict success; target-frame behavior, maximum below-plane violation, trajectory length, and episode duration
Physical benchmark
150 scheduled targets per controller; N=97 common-success targets for continuous comparisons
Target-reaching reliability, endpoint error, local TCP acceleration and jerk, joint-position tracking error, normalized target convergence, and trajectory geometry
Greenhouse harvesting
144 harvesting attempts
Target reaching, cutting, retention, release, and complete harvesting success
Table 1: Evaluation protocol used for simulation, controlled physical benchmarking, and greenhouse harvesting.
Metric
PSA-PPO
w/o path safety
w/o side corridor
w/o motion reg.
Final reach (%)
99.0
89.0
100.0
98.0
Path-safe (%)
92.5
69.5
73.0
88.5
Strict success (%)
92.0
63.5
73.0
87.0
Final TCP error (mm)
32.3 [29.2, 35.1]
32.1 [28.3, 38.3]
25.3 [21.4, 28.6]
35.0 [30.5, 39.2]
P95 maximum below-plane violation (mm)
2.11
26.8
14.3
4.52
Trajectory length (m)
0.406 [0.264, 0.547]
0.443 [0.294, 0.567]
0.461 [0.294, 0.592]
0.464 [0.309, 0.592]
Table 2: Component-removal ablation of PSA-PPO over the same 200 simulation targets. Continuous quantities are reported as median [Q1,Q3].
Figure 5: Effect of PSA-PPO component removal on harvesting-compatible simulation performance over the same 200 targets. (a) Final reaching, path-safe reaching, and strict success with Wilson 95% confidence intervals. (b) Empirical distribution of the maximum wrist–TCP below-plane violation; the dashed vertical line denotes the 2-mm strict path-safety threshold.
Figure 6: Target-frame trajectory profiles for PSA-PPO and the three component-removal formulations over the common 200-target evaluation set: (a) absolute lateral error, (b) absolute vertical error, (c) wrist–TCP below-plane violation, and (d) TCP distance to the target as functions of normalized staging-to-target progress. Curves denote median behavior and shaded regions the corresponding trajectory dispersion.
Figure 7: Strict-success performance of PSA-PPO and the three component-removal formulations across common bins of target radius, target azimuth, and target height (left to right). The bin-wise results characterize variation over the evaluated target set for the selected checkpoints and not variability across independent policy retraining.
Controller
Reached
Failed
Success (%) [95% CI]
PSA-PPO
142/150
8
94.7 [89.8, 97.3]
Cartesian+TOTG
129/150
21
86.0 [79.5, 90.7]
RRTConnect
132/150
18
88.0 [81.8, 92.3]
RRT*
120/150
30
80.0 [72.9, 85.6]
Table 3: Physical target-reaching reliability of PSA-PPO and the classical planning baselines under the common 20-mm reaching criterion. Here, CI denotes the confidence interval.
Figure 8: Empirical cumulative distribution of physical TCP endpoint error. The vertical dashed line denotes the 20-mm reaching criterion. Planner-based control produced tighter endpoint distributions, whereas PSA-PPO exhibited a broader but predominantly tolerance-compliant distribution.
Figure 9: Recorded failure-related events during the physical controller benchmark. Cell counts denote diagnostic events and are not necessarily mutually exclusive; the corresponding numbers of unique unsuccessful targets were 8 for PSA-PPO, 21 for Cartesian+TOTG, 18 for RRTConnect, and 30 for RRT*.
Metric
PSA-PPO
Cartesian+TOTG
RRTConnect
RRT*
RMS acceleration (m/s 2 )
0.1270
0.5277
0.5549
0.2668
RMS jerk (m/s 3 )
11.38
42.17
40.74
18.91
99.5th-percentile jerk (m/s 3 )
77.72
131.88
164.77
127.84
Joint-position tracking RMSE (rad)
6.96×10−4
2.56×10−3
2.47×10−3
1.15×10−3
Table 4: Physical motion quality within the final 0.22-m target-relative approach. Values are medians over the common-success target set ( N=97 ). Lower values indicate lower dynamic excitation or tracking error.
Figure 10: Measured trajectory profiles for the four physical controllers over the common-success target population: (a) TCP speed, (b) acceleration, (c) jerk over normalized active approach progression, and (d) perpendicular deviation from the start–target reference line over normalized longitudinal progress. The normalized axes describe trajectory progression rather than absolute wall-clock execution time.
Figure 11: Normalized target-distance contraction during the final 0.22-m approach over the common-success target set. Remaining TCP-to-target distance is normalized by its value at local-window entry, while the horizontal axis represents each trajectory’s normalized active progression. PSA-PPO reached 50% remaining distance earlier in normalized progression than the three planner-based controllers.
Harvesting stage
Successful
Overall (%)
Conditional (%)
Target reached
136/144
94.4
94.4
Lettuce cut
124/144
86.1
91.2
Table 5: Stage-wise performance of the integrated robotic lettuce harvesting system under greenhouse conditions.
Figure 12: Representative online target stream during greenhouse harvesting. The image was acquired from the robot-mounted RGB-D camera while the UR10e remained at the home configuration. The vision node identified the available lettuce targets and published their corresponding 3D cutting-point coordinates for subsequent robot-frame transformation and target-conditioned reaching.
Figure 13: Representative greenhouse execution of the integrated harvesting pipeline: (a) PSA-PPO-controlled side approach toward the vision-derived cutting location; (b) simultaneous cutting and retention using the 2FG7-driven cut-and-hold end effector; and (c) release of the harvested lettuce into the collection basket.
Figure 14: Physical controller trade-offs between target-reaching reliability and (a) median RMS jerk within the final 0.22-m approach and (b) median trajectory detour ratio. PSA-PPO favors the reliability–dynamic-excitation operating regime, whereas the planner-based controllers retain greater geometric directness.
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Definition
Observation dimension
42
Policy input dimension
Action dimension
6
UR10e joint-space action dimension
Simulation step
1/120 s
Physics integration step
Control decimation
4
Physics steps per policy action
Policy frequency
30 Hz
Effective policy rate
Episode limit
9.0 s
Maximum training episode duration
Appendix
Table A.1 : Policy interaction, action-processing, capture, and path-safety parameters.
Setting
Value
Implementation
Training-target distribution
Target radius
[0.60,1.25] m
Horizontal front-workspace radius
Target azimuth
[−85∘,85∘]
Front-workspace azimuth
Target height
[0.50,0.68] m
Vertical target range
Approach distance da
[0.18,0.28] m
Same-height staging-to-target distance
Base radial sampling
Area-uniform
ρ=ρmin2+u(ρmax2−ρmin2)
Appendix
Table A.2 : Training-workspace sampling and target-conditioned safe-start configuration.
Component
Implemented term
Weight
Definition / activation
Rttar
Staging-distance progress; staging distance
+25;−10
Multiplied by 1−σ(14(ℓt−0.02))
Rttar
Target-distance progress; axial-error progress; line-progress increment; clipped line progress
+720;+610;+360;+185
Distance/axial increments clipped to ±0.04 ; line increment to ±0.05 ; direct ℓt term clipped to [−0.3,1]
Rttar
exp(−7dtT) ; target shells at dtT<0.15,0.10,0.07 m
Alignment reward exp[−5(1−cta)] ; hard/soft vertical thresholds are 0.12/0.085
Appendix
Table A.3 : Implementation-level reward terms used for the final PSA-PPO policy. The component symbols correspond to the compact decomposition in the main text.
Formulation
Training modification relative to PSA-PPO
Preserved elements
PSA-PPO
No component removed. The +2400 strict-capture bonus required final capture, acceptable safe-start clearance, and acceptable accumulated path clearance. Successful termination used the same strict predicate, and accumulated below-plane violation above 0.050 m triggered hard-safety termination.
Complete target, corridor, orientation, path-safety, and motion-regularization formulation.
w/o path safety (C2b)
Removed the five path-safety reward terms: +80I[bt≤εsafe] , −4200bt , −12000bt2 , −3600max(btmax−εsafe,0) , and −2600b0 . The +2400 bonus remained, but its gate was changed from strict capture to final capture only. Training success termination likewise used final capture only, and hard-safety termination was disabled.
The 42-D observation, including wrist–TCP clearance channels, was unchanged. Target/corridor/orientation rewards, motion regularization, safe-start logic, and path-safety logging were retained. Strict path safety was restored only in the common post-training evaluator.
Target-distance and target-frame error terms, orientation, path-safety rewards/history, strict-capture termination, hard-safety termination, and motion regularization were unchanged.
w/o motion regularization (C2d)
Removed the four action/joint regularizers: −0.0035∥ut∥22 , −0.015∥Δut∥22 , −0.026∥Δ2ut∥22 , and −0.00050∥q˙t∥22 . The constant per-step cost −0.10 was retained.
Target, corridor, orientation, path-safety, strict-capture, and termination logic were unchanged.
Appendix
Table A.4 : Exact training modifications used for the PSA-PPO component-removal ablations. Only the listed reward or termination logic was changed relative to the full formulation; observation dimensionality, action processing, target sampling, safe-start construction, capture thresholds, and the post-training strict evaluator were retained.
Parameter
Value
Setting
Training seed
42
RSL-RL seed
Rollout length
32 steps/env.
Samples per environment/update
Maximum iterations
12,000
PPO iterations
Actor hidden layers
256–256–128
Fully connected
Critic hidden layers
256–256–128
Fully connected
Activation
ELU
Hidden activation
Appendix
Table A.5 : PPO optimization and actor–critic network configuration.
Table B.2 : Paired binary outcome differences between PSA-PPO and the three component-removal formulations over the common 200-target set. Positive Δ favors PSA-PPO.
Figure B.1 : Paired strict outcomes for each component removal over the same 200 targets. Each bar partitions targets into cases where both policies were strictly successful, only PSA-PPO was successful, only the ablated policy was successful, or both failed. Δ denotes the strict-success rate of PSA-PPO minus that of the corresponding ablation.
Figure B.2 : Target-wise simulation outcomes over the common workspace for PSA-PPO and the three component-removal formulations. Markers distinguish strict success, safety-only failure, final-only failure, and failure of both criteria. The plots localize evaluation failures but do not represent variation across policy retraining seeds.
Dimension
Interval
N
Final (%)
Path-safe (%)
Strict (%)
Strict 95% CI
Median error / P95 violation (mm)
Radius
[0.6000, 0.7625) m
44
97.7
95.5
93.2
[81.8, 97.7]
33.16 / 1.34
Radius
[0.7625, 0.9250) m
45
100.0
97.8
97.8
[88.4, 99.6]
33.63 / 0.00
Radius
[0.9250, 1.0875) m
47
100.0
95.7
95.7
[85.8, 98.8]
29.89 / 0.36
Radius
[1.0875, 1.2500] m
64
98.4
84.4
84.4
[73.6, 91.3]
33.15 / 3.63
Azimuth
[-85.0, -51.0) deg
37
97.3
78.4
78.4
[62.8, 88.6]
32.40 / 6.03
Azimuth
[-51.0, -17.0) deg
42
100.0
95.2
95.2
[84.2, 98.7]
33.11 / 0.62
Appendix
Table B.3 : Numerical workspace robustness of the selected PSA-PPO checkpoint across predefined equal-width bins.
Figure B.3 : Representative trajectory overlays for the three component-removal contrasts. The displayed targets were selected from the common evaluation set to expose the corresponding contrast: target 65 for path-safety removal, target 31 for side-corridor removal, and target 199 for motion-regularization removal. The star marks the target location. These examples supplement, rather than replace, the paired population statistics.
Figure B.4 : Post-hoc fixed-band corridor diagnostics over the common target set: fraction of trajectory samples inside both lateral and vertical bands, per-target P95 absolute lateral error, and per-target P95 absolute vertical error. The distributions do not exhibit a uniform ordering across formulations and should not be interpreted as a direct measurement of the complete corridor reward.
Figure B.5 : Empirical distributions of direct motion-regularization diagnostics: raw-action-rate RMS, raw-action-acceleration RMS, and joint-velocity-norm RMS. The no-motion-regularization policy does not shift these quantities uniformly toward larger values, demonstrating that the component removal changes the learned operating point rather than providing a monotonic smoothness test.
Figure B.6 : Final tool-orientation distributions over the common evaluation set: approach-axis alignment (left) and absolute tool-axis vertical component (right). Orientation terms and final orientation criteria were unchanged across the three component-removal formulations.
Parameter
Value
Role
HOME TCP position
[−0.164,0.641,0.382] m
Robot observation/home pose
Benchmark tool orientation (w,x,y,z)
[0.009162454,0.005118748,0.718754832,0.695184393]
Common fixed physical orientation
Base frame
robot_arm_base_link
Cartesian reference frame
Tool frame
robot_arm_tool0
Measured TCP frame
Paper evaluation tolerance
0.020 m
Binary target-reaching criterion
ROS monitor tolerance
0.002 m
Online execution-monitor threshold
Appendix
Table C.1 : Principal physical execution and logging parameters used for PSA-PPO deployment and the controlled real-robot benchmark.
Setting
Value
Application
Move group
arm
All MoveIt baselines
RRTConnect backend
OMPL RRTConnect ( RRTConnectkConfigDefault )
Pose planning using the frozen RRTConnect planner ID
RRT* backend
OMPL RRT* ( RRTstar )
Pose planning using the post-fix RRT* planner ID
Cartesian backend
Cartesian linear + TOTG ( CartesianLinearTOTG )
Collision-checked Cartesian interpolation followed by retiming
Planning time
8.0 s
OMPL pose planners
Planning attempts
3
OMPL pose planners
Appendix
Table C.2 : MoveIt baseline configuration used in the controlled physical benchmark. Parameters listed as common were applied to the planner-based controllers unless the backend required a method-specific setting.
PSA-PPO
Cartesian+TOTG
RRTConnect
RRT*
Scheduled targets
150
150
150
150
Archived measured trajectories
150
150
150
125
Paper-level reached targets
142
129
132
120
Paper-level failed targets
8
21
18
30
Appendix
Table C.3 : Physical benchmark populations used for binary and continuous controller comparisons.
Metric
PSA-PPO
Cartesian+TOTG
RRTConnect
RRT*
RMS acceleration (m/s 2 )
0.1776
0.5776
0.5856
0.2739
RMS jerk (m/s 3 )
15.82
47.68
45.01
19.91
99.5th-percentile jerk (m/s 3 )
105.11
143.38
172.29
127.58
Joint-position tracking RMSE (rad)
8.41×10−4
2.87×10−3
2.76×10−3
1.31×10−3
Appendix
Table C.4 : Selected physical motion metrics over the complete reconstructed approach on the common-success target set ( N=97 ). Values are controller medians. Lower values indicate lower dynamic excitation or tracking error.
Metric
Planner
PSA-PPO reduction (%)
Median paired difference [95% CI]
Holm-adjusted p
RMS acceleration (m/s 2 )
Cartesian+TOTG
75.9
−0.4012 [ −0.4118,−0.3953 ]
4.02×10−16
RRTConnect
77.1
−0.4295 [ −0.4419,−0.4162 ]
4.02×10−16
RRT*
52.4
−0.1332 [ −0.1398,−0.1303 ]
4.02×10−16
RMS jerk (m/s 3 )
Cartesian+TOTG
73.0
−30.71 [ −31.79,−29.83 ]
4.02×10−16
RRTConnect
72.1
−29.34 [ −30.62,−28.54 ]
4.02×10−16
RRT*
39.8
−7.177 [ −7.625,−6.442 ]
4.02×10−16
Appendix
Table C.5 : Paired effect-size analysis for PSA-PPO relative to each classical planner within the final 0.22-m target-relative approach. Differences are computed target-wise over the identical common-success set ( N=97 ).
Controller
Median TCP error (mm)
Fraction of 20-mm tolerance (%)
PSA-PPO
9.78
48.9
Cartesian+TOTG
0.06
0.3
RRTConnect
1.62
8.1
RRT*
1.61
8.1
Appendix
Table C.6 : Terminal positioning accuracy on the common-success physical target set. The physical reaching tolerance was 20 mm.
Metric
PSA-PPO
Cartesian+TOTG
RRTConnect
RRT*
Detour ratio
1.8094
1.6697
1.6812
1.6962
Path efficiency
0.5527
0.5989
0.5948
0.5896
Backtracking ratio
0.0371
0.0178
0.0195
0.0218
Total Cartesian turning ( ∘ )
410.09
16.91
75.58
75.14
Base-azimuth travel ( ∘ )
26.25
25.73
25.97
26.45
Total joint travel (rad)
3.108
2.850
2.910
2.723
Appendix
Table C.7 : Trajectory-geometry and spatial-excursion statistics over the full reconstructed physical approach. Median quantities characterize typical path geometry; P95 and maximum deviations characterize distribution tails.
Planner
Witnessed failures
Additional ranked identities selected
Archived trajectories
Hard-event selected trials
Highest-priorityt diagnostic target
Cartesian+TOTG
21
19
150
0
T119
RRTConnect
18
18
150
3
T040
RRT*
30
5
125
4
T087
Appendix
Table C.8 : Summary of the anomaly-ranked witnessed-count sensitivity construction for the planner baselines.
Figure C.1 : Minimum wrist- and elbow-angle margins to the nearest multiple of π in the physical benchmark. The dashed 5∘ reference denotes the diagnostic threshold used by the anomaly-ranking sensitivity analysis. The quantity is kinematic evidence only; near-threshold configurations are not interpreted as proof of singularity-induced failure.
Figure C.2 : Matched target-level trajectory diagnostics from the anomaly-ranked sensitivity analysis: (a) PSA-PPO and Cartesian+TOTG on T119, (b) PSA-PPO and RRTConnect on T040, and (c) PSA-PPO and RRT* on T087. The examples are the highest-priority selected planner cases for their respective methods and are not representative population trajectories. Spatially expansive motion and near-singularity diagnostics are not interpreted as collision events or as proof of failure causation.
This study presents a closed-loop robotic strawberry harvesting system that combines a robust vision module, simulation-trained deep reinforcement learning (DRL) control, and ROS-based realrobot execution. For perception, we propose HRAttnEdge-YOLO26-seg, a modified YOLO26-seg architecture that incorporates a high-resolution P2 branch, segmentation-path attention, and edgesupervised prototype learning to improve instance segmentation in cluttered scenes. For control, we train a target-conditioned Proximal Policy Optimization (PPO) policy in Isaac Lab to produce smooth joint-position commands for a UR10e manipulator and deploy it on a UR10e robot for targetfruit reaching and harvesting. This simulation-based approach reduces hardware dependency, lowers development cost, and allows scalable policy training without exhaustive physical trials before real deployment. The proposed vision model demonstrated the highest overall performance among the evaluated methods. On both self-collected and public datasets, the model showed a 10 to 14% improvement in segmentation performance. In controlled in-house tests, the PPO controller produced stable and dynamically smoother motion than a inverse kinematics (IK)-based MoveIt baseline. In greenhouse trials, the proposed integrated system harvested 281 strawberries, achieving 96.6% reaching success, 91.3% grasp-and-pull success, and 84.3% overall harvesting success. These results illustrate that task-specific perception combined with simulation-trained PPO can serve as a practical and resource-efficient alternative to conventional planner-dependent reaching in manipulation, enabling reliable closed-loop robotic harvesting in complex agricultural environments.
Al Bashir, Shao-Yang Chang, Partho Ghose +3
Department of Biological & Agricultural Engineering, Texas A&M University, College Station, TX 77843, USA · Department of Biomechatronics Engineering, National Taiwan University, 10617, Taipei City, Taiwan · Texas A&M AgriLife Research, Texas A&M University System, Dallas, TX 75252, USA
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.
Teng Li, Hanfei Shi, Chunjiang Zhao +1
The Intelligent Equipment Research Center, Beijing Academy of Agriculture and Forestry Sciences, Beijing 100097, China. · School of Intelligence Science and Technology, University of Science and Technology Beijing, Beijing 100083, China. · Happy Elements Ltd., Beijing 100094, China.
Severe occlusions and deformable plant structures introduce complex contact dynamics that challenge robotic strawberry harvesting. A policy-driven reinforcement learning (RL) framework with heuristic phase coordination was developed, in which obstacle separation, fruit detachment, and placement were formulated as a sequential decision-making task. A shared interaction-aware policy generated Cartesian motions across all task phases, while lightweight heuristic logic coordinated task progression and gripper events. A shared structured observation space was used to represent target, obstacle, end-effector, and task-context information. A hierarchical architecture combined the high-level policy with low-level Cartesian impedance control for compliant interaction. To support zero-shot sim-to-real transfer, feasibility-first observation alignment and domain randomization were adopted. The policy achieved success rates of 89.7% in simulation and 82.0% in real-world experiments. As the occlusion level increased from 1 to 5, the average execution time increased from 12.99 s to 21.73 s, reflecting greater interaction complexity. These results demonstrated effective transfer of interaction-aware harvesting behaviors to a structurally different robotic platform.
Changyou Miao, Teng Li, Ya Xiong
The Intelligent Equipment Research Center, Beijing Academy of Agriculture and Forestry Sciences, Beijing 100097, China. · College of Engineering, Huazhong Agricultural University, Wuhan 430070, China