Cabinet Door Opening Robot

Overview

This repository contains a RoboCasa OpenCabinet project built around two use cases:

A compact starter workflow for exploring the environment, downloading data, and training a simple baseline.
The final project experiments, including the BC U-Net and diffusion U-Net workflows documented below.

The robot

We use the PandaOmron mobile manipulator -- a Franka Panda 7-DOF arm mounted on an Omron wheeled base with a torso lift joint. This is the default and best-supported robot in RoboCasa.

Installation

Run the install script (works on macOS and WSL/Linux):

./install.sh

This will:

Create a Python virtual environment (.venv)
Clone and install robosuite and robocasa
Install all Python dependencies (PyTorch, numpy, matplotlib, etc.)
Download RoboCasa kitchen assets (~10 GB)

After installation, activate the environment:

source .venv/bin/activate

Then verify everything works:

cd cabinet_door_project
python 00_verify_installation.py

macOS note: Scripts that open a rendering window (03, 05) require mjpython instead of python. The install script will remind you of this.

Project Structure

cabinet_door_project/
  00_verify_installation.py
  01_explore_environment.py
  02_random_rollouts.py
  03_teleop_collect_demos.py
  04_download_dataset.py
  05_playback_demonstrations.py
  06_train_policy.py
  07_evaluate_policy.py
  08_visualize_policy_rollout.py
  prepare_dataset.py             # Build oracle / handle feature caches
  validate_best.py               # Reproduce the final pretrain BC U-Net run
  bc_handle.py                   # Handle-aware BC / diffusion training entrypoint
  compare_bc_vs_diffusion.py     # Matched BC vs diffusion experiments
  ablation_sweep.py              # Feature / method sweep code
  preprocess_all_states.py       # Precompute oracle state features
  generate_door_positions.py     # Replay-based door geometry extraction
  diffusion_policy/              # Core training / inference / evaluation package
  tests/                         # Regression tests for the diffusion policy code
  configs/
    diffusion_policy.yaml
install.sh
README.md

Starter Workflow

For the lightweight baseline path:

python 00_verify_installation.py
python 01_explore_environment.py
python 04_download_dataset.py
python 05_playback_demonstrations.py
python 06_train_policy.py
python 07_evaluate_policy.py --checkpoint /tmp/cabinet_policy_checkpoints/best_policy.pt

Use 03_teleop_collect_demos.py if you want to collect demonstrations manually. On macOS, scripts that open a renderer window should be run with mjpython.

Reproducing The Main U-Net Experiments

If you want to reproduce the main U-Net workflows used in this project, use the commands below instead of the starter 06_train_policy.py script. These commands assume you already ran ./install.sh, activated .venv, and are working from cabinet_door_project/.

Step A: Prepare The Oracle / Handle Features

Download the dataset, then build the cached oracle features used by the final U-Net experiments:

python 04_download_dataset.py
python prepare_dataset.py --n_workers 4
python prepare_dataset.py --validate_only

This produces the artifacts consumed by the final experiments under /tmp/diffusion_policy_checkpoints/, including:

door_positions.npz
door_quats.npz
preprocessed_all_states.pt
handle_cache/

Step B: Reproduce The Final Pretrain BC U-Net Result

Run:

python validate_best.py

What this does:

Retrains the final F3 BC U-Net on the 107-demo pretrain dataset
Uses the 22-dim handle-aware state: proprio + handle_pos + handle_to_eef
Evaluates 100 episodes with the relaxed one-door-open success criterion

Expected output:

Checkpoint written to /tmp/diffusion_policy_checkpoints/best_f3_bc_unet.pt
Final summary line close to:

FINAL RESULT: 44/100 success (44.0%)

Wall-clock time depends heavily on GPU / CPU count. Training is short; the 100-episode evaluation is the expensive part.

Step C: Reproduce The Checked-In Diffusion U-Net Run

Run:

python bc_handle.py \
  --arch unet \
  --horizon 16 \
  --n_obs_steps 2 \
  --n_action_steps 8 \
  --epochs 300 \
  --patience 40 \
  --ddpm_steps 100 \
  --ddim_steps 10 \
  --n_eps 20 \
  --max_steps 500

What this does:

Trains the repository's handle-aware diffusion U-Net
Uses the 44-dim handle-relative oracle state from bc_handle.py
Evaluates 20 episodes in the same pretrain-only setting

Outputs:

Checkpoint written to /tmp/diffusion_policy_checkpoints/bc_unet_best.pt
Eval summary printed at the end as:

Result: X/20 (Y.Y%)

To re-run eval only from the saved checkpoint:

python bc_handle.py \
  --arch unet \
  --eval_only \
  --checkpoint /tmp/diffusion_policy_checkpoints/bc_unet_best.pt \
  --n_eps 20 \
  --max_steps 500

Notes

The BC U-Net path above is the reproducible path for the 44/100 pretrain result reported in the final writeup.
The diffusion U-Net command above reproduces the checked-in pretrain diffusion U-Net pipeline in this repo.

Key Concepts

The OpenCabinet Task

Goal: Open a kitchen cabinet door
Fixture: HingeCabinet (a cabinet with hinged doors)
Initial state: Cabinet door is closed; robot is positioned nearby
Success: fixture.is_open(env) returns True
Horizon: 500 timesteps at 20 Hz control frequency (25 seconds)
Scene variety: 2,500+ kitchen layouts/styles for generalization

Observation Space (PandaOmron)

Key	Shape	Description
`robot0_agentview_left_image`	(256, 256, 3)	Left shoulder camera
`robot0_agentview_right_image`	(256, 256, 3)	Right shoulder camera
`robot0_eye_in_hand_image`	(256, 256, 3)	Wrist-mounted camera
`robot0_gripper_qpos`	(2,)	Gripper finger positions
`robot0_base_pos`	(3,)	Base position (x, y, z)
`robot0_base_quat`	(4,)	Base orientation quaternion
`robot0_base_to_eef_pos`	(3,)	End-effector pos relative to base
`robot0_base_to_eef_quat`	(4,)	End-effector orientation relative to base

Action Space (PandaOmron)

Key	Dim	Description
`end_effector_position`	3	Delta (dx, dy, dz) for the end-effector
`end_effector_rotation`	3	Delta rotation (axis-angle)
`gripper_close`	1	0 = open, 1 = close
`base_motion`	4	(forward, side, yaw, torso)
`control_mode`	1	0 = arm control, 1 = base control

Dataset Format (LeRobot)

Datasets are stored in LeRobot format:

dataset/
  meta/           # Episode metadata (task descriptions, camera info)
  videos/         # MP4 videos from each camera
  data/           # Parquet files with actions, states, rewards
  extras/         # Per-episode metadata

Troubleshooting

I'll continually update this section as students find bugs in the system. Please, let me know if you encounter issues!

Problem	Solution
`MuJoCo version must be 3.3.1`	`pip install mujoco==3.3.1`
`numpy version must be 2.2.5`	`pip install numpy==2.2.5`
Rendering crashes on Mac	Use `mjpython` instead of `python`
`GLFW error` on headless server	Set `export MUJOCO_GL=egl` or `osmesa`
Out of GPU memory during training	Reduce batch size in `configs/diffusion_policy.yaml`
Kitchen assets not found	Run `python -m robocasa.scripts.download_kitchen_assets`

ncylich cs188-cabinet-door-project

Cabinet Door Opening Robot

Overview

The robot

Installation

Project Structure

Starter Workflow

Reproducing The Main U-Net Experiments

Step A: Prepare The Oracle / Handle Features

Step B: Reproduce The Final Pretrain BC U-Net Result

Step C: Reproduce The Checked-In Diffusion U-Net Run

Notes

Key Concepts

The OpenCabinet Task

Observation Space (PandaOmron)

Action Space (PandaOmron)

Dataset Format (LeRobot)

Troubleshooting

References

ncylich/cs188-cabinet-door-project

Cabinet Door Opening Robot

Overview

The robot

Installation

Project Structure

Starter Workflow

Reproducing The Main U-Net Experiments

Step A: Prepare The Oracle / Handle Features

Step B: Reproduce The Final Pretrain BC U-Net Result

Step C: Reproduce The Checked-In Diffusion U-Net Run

Notes

Key Concepts

The OpenCabinet Task

Observation Space (PandaOmron)

Action Space (PandaOmron)

Dataset Format (LeRobot)

Troubleshooting

References

ncylich cs188-cabinet-door-project