Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ChicGrasp: Imitation-Learning based Customized Dual-Jaw Gripper Control for Delicate, Irregular Bio-products Manipulation

Amirreza Davar, Zhengtong Xu, Siavash Mahmoudi, Pouya Sohrabipour, Chaitanya Pallerla, Yu She, Wan Shou, Philip Crandall, Dongyi Wang

TL;DR

The paper tackles the problem of automating the handling of delicate, deformable poultry carcasses in processing lines, where suction and rigid, scripted robots struggle. It introduces ChicGrasp, a hardware–software co-design consisting of a custom independently actuated dual-jaw gripper and a conditional diffusion policy trained from 50 multi-view demonstrations to plan $5$-DoF end-effector motion, including jaw commands, bound to RGB and proprioceptive inputs. The system achieves a 40.6% real-world grasp-and-rehang success rate (41/101) across three carcass exemplars, outperforming two baselines (IBC and LSTM-GMM) that fail to complete tasks; the approach is demonstrated with a hybrid learned grasp and scripted rehang. Although promising, the authors note speed limitations and data scarcity, and propose future work in end-to-end learning of the full sequence, richer sensing, and broader applicability to other delicate foods, with open-source CAD, code, and datasets to enable reproducible benchmarking in agrobotics.

Abstract

Automated poultry processing lines still rely on humans to lift slippery, easily bruised carcasses onto a shackle conveyor. Deformability, anatomical variance, and strict hygiene rules make conventional suction and scripted motions unreliable. We present ChicGrasp, an end--to--end hardware--software co-design for this task. An independently actuated dual-jaw pneumatic gripper clamps both chicken legs, while a conditional diffusion-policy controller, trained from only 50 multi--view teleoperation demonstrations (RGB + proprioception), plans 5 DoF end--effector motion, which includes jaw commands in one shot. On individually presented raw broiler carcasses, our system achieves a 40.6\% grasp--and--lift success rate and completes the pick to shackle cycle in 38 s, whereas state--of--the--art implicit behaviour cloning (IBC) and LSTM-GMM baselines fail entirely. All CAD, code, and datasets will be open-source. ChicGrasp shows that imitation learning can bridge the gap between rigid hardware and variable bio--products, offering a reproducible benchmark and a public dataset for researchers in agricultural engineering and robot learning.

ChicGrasp: Imitation-Learning based Customized Dual-Jaw Gripper Control for Delicate, Irregular Bio-products Manipulation

TL;DR

The paper tackles the problem of automating the handling of delicate, deformable poultry carcasses in processing lines, where suction and rigid, scripted robots struggle. It introduces ChicGrasp, a hardware–software co-design consisting of a custom independently actuated dual-jaw gripper and a conditional diffusion policy trained from 50 multi-view demonstrations to plan -DoF end-effector motion, including jaw commands, bound to RGB and proprioceptive inputs. The system achieves a 40.6% real-world grasp-and-rehang success rate (41/101) across three carcass exemplars, outperforming two baselines (IBC and LSTM-GMM) that fail to complete tasks; the approach is demonstrated with a hybrid learned grasp and scripted rehang. Although promising, the authors note speed limitations and data scarcity, and propose future work in end-to-end learning of the full sequence, richer sensing, and broader applicability to other delicate foods, with open-source CAD, code, and datasets to enable reproducible benchmarking in agrobotics.

Abstract

Automated poultry processing lines still rely on humans to lift slippery, easily bruised carcasses onto a shackle conveyor. Deformability, anatomical variance, and strict hygiene rules make conventional suction and scripted motions unreliable. We present ChicGrasp, an end--to--end hardware--software co-design for this task. An independently actuated dual-jaw pneumatic gripper clamps both chicken legs, while a conditional diffusion-policy controller, trained from only 50 multi--view teleoperation demonstrations (RGB + proprioception), plans 5 DoF end--effector motion, which includes jaw commands in one shot. On individually presented raw broiler carcasses, our system achieves a 40.6\% grasp--and--lift success rate and completes the pick to shackle cycle in 38 s, whereas state--of--the--art implicit behaviour cloning (IBC) and LSTM-GMM baselines fail entirely. All CAD, code, and datasets will be open-source. ChicGrasp shows that imitation learning can bridge the gap between rigid hardware and variable bio--products, offering a reproducible benchmark and a public dataset for researchers in agricultural engineering and robot learning.
Paper Structure (11 sections, 4 equations, 9 figures, 2 tables)

This paper contains 11 sections, 4 equations, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Customized End-Effector Design
  4. Action Space and Policy Formulation
  5. Demonstration Collection
  6. Policy Training
  7. Runtime Control Loop
  8. Results
  9. Discussion
  10. Conclusion
  11. Acknowledgement

Figures (9)

  • Figure 1: (Left) The robotic gripper equipped with two jaws and wrist camera designed for autonomous poultry handling. (Right) Comparison between the current industry-standard manual process and our autonomous approach using imitation learning to pick up and rehang poultry.
  • Figure 2: Poultry processing workflow from receiving to packaging which highlights the targeted automation in the chilling and rehang stage.
  • Figure 3: Learning a grasping policy from demonstration: The network inputs stacked image frames, robot state vectors, and gripper states, and outputs both the end‐effector pose trajectory and gripper control commands.
  • Figure 4: Schematic of the 2DoF customized gripper system for chicken carcass grasping. The system consists of: (1) a 3D-printed interface for robot attachment, (2) a pneumatic actuator for jaw control, (3) 3D-printed serrated jaw surface (30° chevron ridges), (4) a pneumatic valve regulating air input and output, and (5) a control section incorporating a microcontroller for system operation. Air supply is routed through the valve to actuate the pneumatic gripper, enabling controlled gripping and releasing motions.
  • Figure 5: Experimental setup for multi-view data collection. Three RGB cameras are positioned to capture different perspectives of the manipulation space: Camera 1 is mounted directly on the gripper (eye-in-hand configuration) to provide a dynamic front view of the target object, and Camera 2 and Camera 3 are placed on the right and left side of the table, respectively, for a side-overview perspective.
  • ...and 4 more figures