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

ICRA 2026

Zhaodong Yang¹, Ai-Ping Hu^1,2, Harish Ravichandar^1
1Georgia Institute of Technology ²Georgia Tech Research Institute

              

Installation

conda create -n deformable python=3.9.19
conda activate deformable
pip install matplotlib
pip install scipy
pip install taichi
pip install stable-baselines3

For hardware deploy on UR5 robot:

pip install math3d ikpy imutils pyserial bitstring pymodbus pyrealsense2 opencv-python

Residual Policy Training in Simulation

To run the simulation without residual policy:

python -m mpm3d.mpm3d_gym_wrapped.py

Run the residual policy training:

python train_residual.py

To train with specific training steps and without rendering:

python train_residual.py --training_steps=50000 --headless

To evaluate the trained model:

python train_residual.py --eval --model_dir=./replace/with/your/checkpoint/directory

2D MPM-CPIC Interactive Example

python -m mpm2d.mpm2d.py

Drag the 2D "knife" with your mouse to cut through the material: