Towards Automated Chicken Deboning via Learning-based Dynamically-Adaptive 6-DoF Multi-Material Cutting

TL;DR

The paper tackles automated chicken shoulder deboning by enabling precise $6$-DoF knife cutting through occluded, deformable, multi-material joints using a residual RL policy trained in a custom simulator. It delivers three core contributions: a first open-source multi-material deformable cutting simulator with coupling, fracture, and force feedback; a reusable physical testbed that mimics the chicken shoulder; and a learning-based residual policy that adapts nominal trajectories via force information to achieve safe, high-DoF cutting with zero-shot sim-to-real transfer. Across simulation, a simplified real-world model, and real chicken shoulders, the approach substantially reduces bone contact and improves success rates, outperforming open-loop baselines by up to about fourfold and highlighting the critical role of force feedback in multi-material cutting. These results suggest a viable path toward scalable, safe automated deboning and broader high-DOF multi-material cutting tasks in industry.

Abstract

Automating chicken shoulder deboning requires precise 6-DoF cutting through a partially occluded, deformable, multi-material joint, since contact with the bones presents serious health and safety risks. Our work makes both systems-level and algorithmic contributions to train and deploy a reactive force-feedback cutting policy that dynamically adapts a nominal trajectory and enables full 6-DoF knife control to traverse the narrow joint gap while avoiding contact with the bones. First, we introduce an open-source custom-built simulator for multi-material cutting that models coupling, fracture, and cutting forces, and supports reinforcement learning, enabling efficient training and rapid prototyping. Second, we design a reusable physical testbed to emulate the chicken shoulder: two rigid "bone" spheres with controllable pose embedded in a softer block, enabling rigorous and repeatable evaluation while preserving essential multi-material characteristics of the target problem. Third, we train and deploy a residual RL policy, with discretized force observations and domain randomization, enabling robust zero-shot sim-to-real transfer and the first demonstration of a learned policy that debones a real chicken shoulder. Our experiments in our simulator, on our physical testbed, and on real chicken shoulders show that our learned policy reliably navigates the joint gap and reduces undesired bone/cartilage contact, resulting in up to a 4x improvement over existing open-loop cutting baselines in terms of success rate and bone avoidance. Our results also illustrate the necessity of force feedback for safe and effective multi-material cutting. The project website is at https://sites.google.com/view/chickendeboning-2026.

Figures (8)

• Figure 1: Overview of this paper. A. MRI scan of a chicken. The red line indicates the shoulder joint position and orientation, where we need to precisely cut through. B. We first train the residual policy in the simulator. For each time step, the residual policy observe the current knife pose $\xi_t$ and knife force sensor reading $f_t$ from the simulator. It also observes residual pose $e^\xi_{t}$ from the previous time step, and one step future knife pose $\hat{\xi}_{t+1}$ retrieved from the nominal trajectory. The residual policy eventually produces the next residual pose: $e^\xi_{t+1}$. And $\tilde{\xi}_{t+1} = \hat{\xi}_{t+1} + e^\xi_{t+1}$ is the adjusted pose of the nominal trajectory, also the next knife target pose. We use different nominal trajectories for different cutting tasks. Nominal trajectory is an interpolated trajectory of several keyframe poses, which are either fixed or mapped from the bone position. C. We deploy the residual policy on two challenging real-world multi-material cutting tasks.
• Figure 2: The setup process of our simplified reusable physical testbed.
• Figure 3: Roboninja's adaptive cutting policy qualitative evaluation results. The left picture shows that RoboNinja relies on knife-bone collision for interactive state estimation, which does crucial damage to non-rigid core in our task. The right is a failure case of RoboNinja when the knife first cuts into the right bone. Because RoboNinja can only adjust its trajectory to one direction (right), it keeps cutting into the right bone deeper and deeper.
• Figure 4: We use a cone which is clamped on the table to fix the chicken while conducting the cutting experiment.
• Figure 5: Simulation qualitative evaluation result. The first and second rows show typical failure cases of the Nominal and Adaptive w/o Force. Due to lack of adaptivity, Nominal are unable to avoid the bone in most of the cases. Though Adaptive w/o Force has the ability to adjust its trajectory, it can not perform well because it lacks necessary force feedback. The third row demonstrates a typical successful example of Adaptive, which has both the ability of adaptation and receiving necessary force feedback.