GenDP: 3D Semantic Fields for Category-Level Generalizable Diffusion Policy

Website | Paper | Colab | Video

Yixuan Wang¹, Guang Yin², Binghao Huang¹, Tarik Kelestemur³, Jiuguang Wang³, Yunzhu Li¹

¹Columbia University, ²University of Illinois Urbana-Champaign, ³Boston Dynamics AI Institute

https://github.com/WangYixuan12/gendp/assets/32333199/f86c977b-bcce-45cf-b632-95663abf3607

:bookmark_tabs: Table of Contents

• Install
• Generate Dataset
  • Generate from Existing Environments
  • Generate from Customized Environments
  • Generate Large-Scale Data
• Download Dataset
• Visualize Dataset
  • Visualize 2D Observation
  • Visualize Aggregated 3D Observation
  • Visualize 3D Semantic Fields
• Train
  • Train in Simulation
  • Config Explanation
• Infer in Simulation
• Deploy in Real World
  • Hardware Prerequisites
  • Install Environment for Real World
  • Set Up Robot
  • Calibrate Camera and Robot Transformation
  • Collect Demonstration
  • Train in Real World
  • Infer in Real World
  • Adapt to New Task

:hammer: Install

We recommend Mambaforge instead of the standard anaconda distribution for faster installation:

mamba env create -f conda_environment.yaml
conda activate gendp
pip install -e gendp/
pip install -e sapien_env/
pip install -e robomimic/
pip install -e d3fields_dev/

:floppy_disk: Generate Dataset

Generate from Existing Environments

We use the SAPIEN to build the simulation environments. To create the data of heuristic policy for single episode, use the following command:

python gen_single_episode.py [episode_idx] [dataset_dir] [task_name] --headless --obj_name [OBJ_NAME] --mode [MODE_NAME]

For example, to generate one episode for the hang_mug task with the GUI, you could run the following command:

python gen_single_episode.py 0 data/ hang_mug --obj_name nescafe_mug # random seed is 0; save the data into data/; task name is hang_mug; object name is nescafe_mug

Meanings for each argument are visible when running python gen_single_episode.py --help.

Generate from Customized Environments

If you want to create your own environments with different objects, please imitate sapien_env/sapien_env/sim_env/mug_collect_env.py. Note that sim_env/custom_env.py does NOT contain the robot. To add robots, please imitate sapien_env/sapien_env/rl_env/mug_collect_env.py to add robots. To adjust camera views, please change YX_TABLE_TOP_CAMERAS within sapien_env/sapien_env/gui/gui_base.py.

Generate Large-Scale Data

We notice that sapien renderer have memory leak for large-scale data generation. To avoid this, we use bash commands to generate large-scale data.

python gen_multi_episodes.py

Arguments can be edited within gen_multi_episodes.py.

:inbox_tray: Download Dataset

If you want to download a small dataset to test the whole pipeline, you can run bash scripts/download_small_data.sh. For hangning mug and pencil insertion task, you can run the following commands:

bash scripts/download_hang_mug.sh
bash scripts/download_pencil_insertion.sh

If the scripts do not work, you could manully download the data from UIUC Box or Google Drive and unzip them.

:art: Visualize Dataset

Visualize 2D Observation

To visualize image observations within hdf5 files, use the following command:

python gendp/tests/vis_data_2d.py 

You could adjust dataset path and observation keys in gendp/tests/vis_data_2d.py.

Visualize Aggregated 3D Observation

Similarly, to visualize aggegated 3D observations, use the following command:

python gendp/tests/vis_aggr_data_3d.py

This will visualize aggregated point clouds from multiple views, robot states, and actions from the dataset. You could adjust dataset path and observation keys in gendp/tests/vis_aggr_data_3d.py.

Visualize 3D Semantic Fields

Similarly, to visualize 3D semantic fields, use the following command:

python gendp/tests/vis_semantic_fields.py

This will visualize 3D semantic fields processed by D3Fields, robot states, and actions. You could adjust dataset path and observation keys in gendp/tests/vis_semantic_fields.py. The explanation of each entries within shape_meta can be seen at Config Explanation.

:gear: Train

Train in Simulation

To run training, we first set the environment variables.

export OMP_NUM_THREADS=1
export TOKENIZERS_PARALLELISM=true
export MKL_NUM_THREADS=1

Then, we run the following command:

cd [PATH_TO_REPO]/gendp
python train.py --config-dir=config/[TASK_NAME] --config-name=distilled_dino_N_4000.yaml training.seed=42 training.device=cuda training.device_id=0 data_root=[PATH_TO_DATA]

For example, to train on small_data in my local machine, I run the following command:

python train.py --config-dir=config/small_data --config-name=distilled_dino_N_4000.yaml training.seed=42 training.device=cuda training.device_id=0 data_root=/home/yixuan/gendp

Please wait at least till 2 epoches to make sure that all pipelines are working properly. For hang_mug_sim task and pencil_insertion_sim task, you could simply replace [TASK_NAME] with hang_mug_sim and pencil_insertion_sim respectively.

Config Explanation

There are several critical entries within the config file. Here are some explanations:

shape_meta: shape_meta contains the policy input and output information.
    action: output information
        shape: action dimension. In our work, it is 10 = (3 for translation, 6 for 6d rotation*, 1 for gripper)
        key: [optional] key for the action in the dataset. It could be 'eef_action' or 'joint_action'. Default is 'eef_action'.
    obs: input information
        ... # other input modalities if needed
        d3fields: 3D semantic fields
            shape: shape of the 3D semantic fields, i.e. (num_channel, num_points)
            type: type of inputs. It should be 'spatial' for point cloud inputs
            info: information of the 3D semantic fields.
                reference_frame: frame of input semantic fields. It should be 'world' or 'robot'
                distill_dino: whether to add semantic information to the point cloud
                distill_obj: the name for reference features, which are saved in `d3fields_dev/d3fields/sel_feats/[DISTILL_OBJ].npy`.
                view_keys: viewpoint keys for the semantic fields.
                N_gripper: number of points sampled from the gripper.
                boundaries: boundaries for the workspace.
                resize_ratio: our pipeline will resize images by this ratio to save time and memory.
task:
    env_runner: the configuration for the evaluation environment during the training
        max_steps: maximum steps for each episode, which should be adjusted according to the task
        n_test: number of testing environments
        n_test_vis: number of testing environments that will be visualized on wandb
        n_train: number of training environments
        n_train_vis: number of training environments that will be visualized on wandb
        train_obj_ls: list of objects that appear in the training environments
        test_obj_ls: list of objects that appear in the testing environments
training:
    checkpoint_every: the frequency of saving checkpoints
    rollout_every: the frequency of rolling out the policy in the env_runner

Also, the configuration might be repetitive in the config file. Please sync them manually.

:video_game: Infer in Simulation

To run an existing policy in the simulator, use the following command:

cd [PATH_TO_REPO]/gendp
python eval.py --checkpoint [PATH_TO_CHECKPOINT] --output_dir [OUTPUT_DIR] --n_test [NUM_TEST] --n_train [NUM_TRAIN] --n_test_vis [NUM_TEST_VIS] --n_train_vis [NUM_TRAIN_VIS] --test_obj_ls [OBJ_NAME_1] --test_obj_ls [OBJ_NAME_2] --data_root [PATH_TO_DATA]

For example, we can run

python eval.py --checkpoint /home/yixuan/gendp/checkpoints/small_data/distilled_dino_N_4000/ckpt_00000000.pt --output_dir /home/yixuan/gendp/eval_results/small_data --n_test 10 --n_train 10 --n_test_vis 5 --n_train_vis 5 --test_obj_ls nescafe_mug --data_root /home/yixuan/gendp

To download the existing checkpoints, you could run the following commands.

bash scripts/download_ckpts.sh

You can also download them from UIUC Box or Google Drive and unzip them if the scipt fails.

:robot: Deploy in Real World

Hardware Prerequisites

• Aloha
• >=1 Realsense Camera

Install Environment for Real World

mamba env create -f conda_environment_real.yaml
pip install -e gendp/
pip install -e d3fields_dev/

Set Up Robot

1. If you already have ROS noetic installed, you could run bash scripts/setup_aloha.sh outside of conda environments. Remember to put source /opt/ros/noetic/setup.sh && source ~/interbotix_ws/devel/setup.sh into ~/.bashrc after installation.
2. As mentioned in Aloha README, you need to go to ~/interbotix_ws/src/interbotix_ros_toolboxes/interbotix_xs_toolbox/interbotix_xs_modules/src/interbotix_xs_modules/arm.py, find function publish_positions. Change self.T_sb = mr.FKinSpace(self.robot_des.M, self.robot_des.Slist, self.joint_commands) to self.T_sb = None. This prevents the code from calculating FK at every step which delays teleoperation.
3. We also need to update usb rules for the robot. You could run the following commands to update the usb rules. You might need to change the serial numbers to your own.

sudo bash scripts/modify_usb_rules.sh
sudo udevadm control --reload && sudo udevadm trigger

1. Remember to reboot the computer after the installation. If you encounter any problems, please refer to the Aloha.
2. To test whether the robot installation is successful, you could run the following command:

# boths sides
roslaunch aloha 4arms_teleop.launch
python gendp/gendp/real_world/aloha_simple_teleop.py --left --right

# left side
roslaunch aloha 2arms_left_teleop.launch
python gendp/gendp/real_world/aloha_simple_teleop.py --left

# right side
roslaunch aloha 2arms_right_teleop.launch
python gendp/gendp/real_world/aloha_simple_teleop.py --right

Calibrate Camera and Robot Transformation

We found raw RealSense intrinsics are accurate enough for our pipeline, but you might want to verify it before proceeding.

First, we calibrate the extrinsics between the camera and the world (i.e. calibration board) frame. We use calib.io to generate the calibration board. Please use ChArUco as Target Type. You could select the rest of options according to your preference and printing capability. Then you can click Save calibration board as PDF to download and print the calibration board. Then you could run

python gendp/gendp/real_world/calibrate_realsenses.py --rows [NUM_ROWS] --cols [NUM_COLS] --checker_width [CHECKER_WIDTH] --marker_width [MARKER_WIDTH]

This will keep running calibration pipeline in a while True loop and save the calibration results in gendp/gendp/real_world/cam_extrinsics. To visualize the calibration results, you could run

python gendp/gendp/real_world/vis_cam_cali.py --iterative

Enabling --iterative will visualize each camera's point cloud iteratively and aggregated point cloud at the end. Otherwise, it will only visualize the aggregated point cloud. You are expected to see a well-aligned point cloud of the workspace.

Lastly, we calibrate the transformations between the robot base and the world frame, which is done manually. You could adjust robot_base_in_world within gendp/gendp/real_world/calibrate_robot.py, which represents the robots' base pose in the world (i.e. calibration board) frame. You could run

python gendp/gendp/real_world/calibrate_robot.py