real-stanford/diffusion_policy

[Project page] [Paper] [Data] [Colab (state)] [Colab (vision)]

Cheng Chi¹, Siyuan Feng², Yilun Du³, Zhenjia Xu¹, Eric Cousineau², Benjamin Burchfiel², Shuran Song¹

¹Columbia University, ²Toyota Research Institute, ³MIT

Our self-contained Google Colab notebooks is the easiest way to play with Diffusion Policy. We provide separate notebooks for state-based environment and vision-based environment.

For each experiment used to generate Table I,II and IV in the paper, we provide:

1. A config.yaml that contains all parameters needed to reproduce the experiment.
2. Detailed training/eval logs.json.txt for every training step.
3. Checkpoints for the best epoch=*-test_mean_score=*.ckpt and last latest.ckpt epoch of each run.

Experiment logs are hosted on our website as nested directories in format: https://diffusion-policy.cs.columbia.edu/data/experiments/<image|low_dim>/<task>/<method>/

Within each experiment directory you may find:

.
├── config.yaml
├── metrics
│   └── logs.json.txt
├── train_0
│   ├── checkpoints
│   │   ├── epoch=0300-test_mean_score=1.000.ckpt
│   │   └── latest.ckpt
│   └── logs.json.txt
├── train_1
│   ├── checkpoints
│   │   ├── epoch=0250-test_mean_score=1.000.ckpt
│   │   └── latest.ckpt
│   └── logs.json.txt
└── train_2
    ├── checkpoints
    │   ├── epoch=0250-test_mean_score=1.000.ckpt
    │   └── latest.ckpt
    └── logs.json.txt

The metrics/logs.json.txt file aggregates evaluation metrics from all 3 training runs every 50 epochs using multirun_metrics.py. The numbers reported in the paper correspond to max and k_min_train_loss aggregation keys.

To download all files in a subdirectory, use:

$ wget --recursive --no-parent --no-host-directories --relative --reject="index.html*" https://diffusion-policy.cs.columbia.edu/data/experiments/low_dim/square_ph/diffusion_policy_cnn/

To reproduce our simulation benchmark results, install our conda environment on a Linux machine with Nvidia GPU. On Ubuntu 20.04 you need to install the following apt packages for mujoco:

$ sudo apt install -y libosmesa6-dev libgl1-mesa-glx libglfw3 patchelf

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

$ mamba env create -f conda_environment.yaml

but you can use conda as well:

$ conda env create -f conda_environment.yaml

The conda_environment_macos.yaml file is only for development on MacOS and does not have full support for benchmarks.

Hardware (for Push-T):

• 1x UR5-CB3 or UR5e (RTDE Interface is required)
• 2x RealSense D415
• 1x 3Dconnexion SpaceMouse (for teleop)
• 1x Millibar Robotics Manual Tool Changer (only need robot side)
• 1x 3D printed End effector
• 1x 3D printed T-block
• USB-C cables and screws for RealSense

Software:

• Ubuntu 20.04.3 (tested)
• Mujoco dependencies: sudo apt install libosmesa6-dev libgl1-mesa-glx libglfw3 patchelf
• RealSense SDK
• Spacemouse dependencies: sudo apt install libspnav-dev spacenavd; sudo systemctl start spacenavd
• Conda environment mamba env create -f conda_environment_real.yaml

Under the repo root, create data subdirectory:

[diffusion_policy]$ mkdir data && cd data

Download the corresponding zip file from https://diffusion-policy.cs.columbia.edu/data/training/

[data]$ wget https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip

Extract training data:

[data]$ unzip pusht.zip && rm -f pusht.zip && cd ..

Grab config file for the corresponding experiment:

[diffusion_policy]$ wget -O image_pusht_diffusion_policy_cnn.yaml https://diffusion-policy.cs.columbia.edu/data/experiments/image/pusht/diffusion_policy_cnn/config.yaml

Activate conda environment and login to wandb (if you haven't already).

[diffusion_policy]$ conda activate robodiff
(robodiff)[diffusion_policy]$ wandb login

Launch training with seed 42 on GPU 0.

(robodiff)[diffusion_policy]$ python train.py --config-dir=. --config-name=image_pusht_diffusion_policy_cnn.yaml training.seed=42 training.device=cuda:0 hydra.run.dir='data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}'

This will create a directory in format data/outputs/yyyy.mm.dd/hh.mm.ss_<method_name>_<task_name> where configs, logs and checkpoints are written to. The policy will be evaluated every 50 epochs with the success rate logged as test/mean_score on wandb, as well as videos for some rollouts.

(robodiff)[diffusion_policy]$ tree data/outputs/2023.03.01/20.02.03_train_diffusion_unet_hybrid_pusht_image -I wandb
data/outputs/2023.03.01/20.02.03_train_diffusion_unet_hybrid_pusht_image
├── checkpoints
│   ├── epoch=0000-test_mean_score=0.134.ckpt
│   └── latest.ckpt
├── .hydra
│   ├── config.yaml
│   ├── hydra.yaml
│   └── overrides.yaml
├── logs.json.txt
├── media
│   ├── 2k5u6wli.mp4
│   ├── 2kvovxms.mp4
│   ├── 2pxd9f6b.mp4
│   ├── 2q5gjt5f.mp4
│   ├── 2sawbf6m.mp4
│   └── 538ubl79.mp4
└── train.log

3 directories, 13 files

Launch local ray cluster. For large scale experiments, you might want to setup an AWS cluster with autoscaling. All other commands remain the same.

(robodiff)[diffusion_policy]$ export CUDA_VISIBLE_DEVICES=0,1,2  # select GPUs to be managed by the ray cluster
(robodiff)[diffusion_policy]$ ray start --head --num-gpus=3

Launch a ray client which will start 3 training workers (3 seeds) and 1 metrics monitor worker.

(robodiff)[diffusion_policy]$ python ray_train_multirun.py --config-dir=. --config-name=image_pusht_diffusion_policy_cnn.yaml --seeds=42,43,44 --monitor_key=test/mean_score -- multi_run.run_dir='data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}' multi_run.wandb_name_base='${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}'

In addition to the wandb log written by each training worker individually, the metrics monitor worker will log to wandb project diffusion_policy_metrics for the metrics aggregated from all 3 training runs. Local config, logs and checkpoints will be written to data/outputs/yyyy.mm.dd/hh.mm.ss_<method_name>_<task_name> in a directory structure identical to our training logs:

(robodiff)[diffusion_policy]$ tree data/outputs/2023.03.01/22.13.58_train_diffusion_unet_hybrid_pusht_image -I 'wandb|media'
data/outputs/2023.03.01/22.13.58_train_diffusion_unet_hybrid_pusht_image
├── config.yaml
├── metrics
│   ├── logs.json.txt
│   ├── metrics.json
│   └── metrics.log
├── train_0
│   ├── checkpoints
│   │   ├── epoch=0000-test_mean_score=0.174.ckpt
│   │   └── latest.ckpt
│   ├── logs.json.txt
│   └── train.log
├── train_1
│   ├── checkpoints
│   │   ├── epoch=0000-test_mean_score=0.131.ckpt
│   │   └── latest.ckpt
│   ├── logs.json.txt
│   └── train.log
└── train_2
    ├── checkpoints
    │   ├── epoch=0000-test_mean_score=0.105.ckpt
    │   └── latest.ckpt
    ├── logs.json.txt
    └── train.log

7 directories, 16 files

Download a checkpoint from the published training log folders, such as https://diffusion-policy.cs.columbia.edu/data/experiments/low_dim/pusht/diffusion_policy_cnn/train_0/checkpoints/epoch=0550-test_mean_score=0.969.ckpt.

Run the evaluation script:

(robodiff)[diffusion_policy]$ python eval.py --checkpoint data/0550-test_mean_score=0.969.ckpt --output_dir data/pusht_eval_output --device cuda:0

This will generate the following directory structure:

(robodiff)[diffusion_policy]$ tree data/pusht_eval_output
data/pusht_eval_output
├── eval_log.json
└── media
    ├── 1fxtno84.mp4
    ├── 224l7jqd.mp4
    ├── 2fo4btlf.mp4
    ├── 2in4cn7a.mp4
    ├── 34b3o2qq.mp4
    └── 3p7jqn32.mp4

1 directory, 7 files

eval_log.json contains metrics that is logged to wandb during training:

(robodiff)[diffusion_policy]$ cat data/pusht_eval_output/eval_log.json
{
  "test/mean_score": 0.9150393806777066,
  "test/sim_max_reward_4300000": 1.0,
  "test/sim_max_reward_4300001": 0.9872969750774386,
...
  "train/sim_video_1": "data/pusht_eval_output//media/2fo4btlf.mp4"
}

Make sure your UR5 robot is running and accepting command from its network interface (emergency stop button within reach at all time), your RealSense cameras plugged in to your workstation (tested with realsense-viewer) and your SpaceMouse connected with the spacenavd daemon running (verify with systemctl status spacenavd).

Start the demonstration collection script. Press "C" to start recording. Use SpaceMouse to move the robot. Press "S" to stop recording.

(robodiff)[diffusion_policy]$ python demo_real_robot.py -o data/demo_pusht_real --robot_ip 192.168.0.204

This should result in a demonstration dataset in data/demo_pusht_real with in the same structure as our example real Push-T training dataset.

To train a Diffusion Policy, launch training with config:

(robodiff)[diffusion_policy]$ python train.py --config-name=train_diffusion_unet_real_image_workspace task.dataset_path=data/demo_pusht_real

Edit diffusion_policy/config/task/real_pusht_image.yaml if your camera setup is different.

Assuming the training has finished and you have a checkpoint at data/outputs/blah/checkpoints/latest.ckpt, launch the evaluation script with:

python eval_real_robot.py -i data/outputs/blah/checkpoints/latest.ckpt -o data/eval_pusht_real --robot_ip 192.168.0.204

Press "C" to start evaluation (handing control over to the policy). Press "S" to stop the current episode.

This codebase is structured under the requirement that:

1. implementing N tasks and M methods will only require O(N+M) amount of code instead of O(N*M)
2. while retaining maximum flexibility.

To achieve this requirement, we

1. maintained a simple unified interface between tasks and methods and
2. made the implementation of the tasks and the methods independent of each other.

These design decisions come at the cost of code repetition between the tasks and the methods. However, we believe that the benefit of being able to add/modify task/methods without affecting the remainder and being able understand a task/method by reading the code linearly outweighs the cost of copying and pasting 😊.

On the task side, we have:

• Dataset: adapts a (third-party) dataset to the interface.
• EnvRunner: executes a Policy that accepts the interface and produce logs and metrics.
• config/task/<task_name>.yaml: contains all information needed to construct Dataset and EnvRunner.
• (optional) Env: an gym==0.21.0 compatible class that encapsulates the task environment.

On the policy side, we have:

• Policy: implements inference according to the interface and part of the training process.
• Workspace: manages the life-cycle of training and evaluation (interleaved) of a method.
• config/<workspace_name>.yaml: contains all information needed to construct Policy and Workspace.

A LowdimPolicy takes observation dictionary:

• "obs": Tensor of shape (B,To,Do)

and predicts action dictionary:

• "action": Tensor of shape (B,Ta,Da)

A LowdimDataset returns a sample of dictionary:

• "obs": Tensor of shape (To, Do)
• "action": Tensor of shape (Ta, Da)

Its get_normalizer method returns a LinearNormalizer with keys "obs","action".

The Policy handles normalization on GPU with its copy of the LinearNormalizer. The parameters of the LinearNormalizer is saved as part of the Policy's weights checkpoint.

A ImagePolicy takes observation dictionary:

• "key0": Tensor of shape (B,To,*)
• "key1": Tensor of shape e.g. (B,To,H,W,3) ([0,1] float32)

and predicts action dictionary:

• "action": Tensor of shape (B,Ta,Da)

A ImageDataset returns a sample of dictionary:

• "obs": Dict of
  • "key0": Tensor of shape (To, *)
  • "key1": Tensor fo shape (To,H,W,3)
• "action": Tensor of shape (Ta, Da)

Its get_normalizer method returns a LinearNormalizer with keys "key0","key1","action".

To = 3
Ta = 4
T = 6
|o|o|o|
| | |a|a|a|a|
|o|o|
| |a|a|a|a|a|
| | | | |a|a|

Terminology in the paper: varname in the codebase

• Observation Horizon: To|n_obs_steps
• Action Horizon: Ta|n_action_steps
• Prediction Horizon: T|horizon

The classical (e.g. MDP) single step observation/action formulation is included as a special case where To=1 and Ta=1.

A Workspace object encapsulates all states and code needed to run an experiment.

• Inherits from BaseWorkspace.
• A single OmegaConf config object generated by hydra should contain all information needed to construct the Workspace object and running experiments. This config correspond to config/<workspace_name>.yaml + hydra overrides.
• The run method contains the entire pipeline for the experiment.
• Checkpoints happen at the Workspace level. All training states implemented as object attributes are automatically saved by the save_checkpoint method.
• All other states for the experiment should be implemented as local variables in the run method.

The entrypoint for training is train.py which uses @hydra.main decorator. Read hydra's official documentation for command line arguments and config overrides. For example, the argument task=<task_name> will replace the task subtree of the config with the content of config/task/<task_name>.yaml, thereby selecting the task to run for this experiment.

A Dataset object:

• Inherits from torch.utils.data.Dataset.
• Returns a sample conforming to the interface depending on whether the task has Low Dim or Image observations.
• Has a method get_normalizer that returns a LinearNormalizer conforming to the interface.

Normalization is a very common source of bugs during project development. It is sometimes helpful to print out the specific scale and bias vectors used for each key in the LinearNormalizer.

Most of our implementations of Dataset uses a combination of ReplayBuffer and SequenceSampler to generate samples. Correctly handling padding at the beginning and the end of each demonstration episode according to To and Ta is important for good performance. Please read our SequenceSampler before implementing your own sampling method.

A Policy object:

• Inherits from BaseLowdimPolicy or BaseImagePolicy.
• Has a method predict_action that given observation dict, predicts actions conforming to the interface.
• Has a method set_normalizer that takes in a LinearNormalizer and handles observation/action normalization internally in the policy.
• (optional) Might has a method compute_loss that takes in a batch and returns the loss to be optimized.
• (optional) Usually each Policy class correspond to a Workspace class due to the differences of training and evaluation process between methods.

A EnvRunner object abstracts away the subtle differences between different task environments.

• Has a method run that takes a Policy object for evaluation, and returns a dict of logs and metrics. Each value should be compatible with wandb.log.

To maximize evaluation speed, we usually vectorize environments using our modification of gym.vector.AsyncVectorEnv which runs each individual environment in a separate process (workaround python GIL).

⚠️ Since subprocesses are launched using fork on linux, you need to be specially careful for environments that creates its OpenGL context during initialization (e.g. robosuite) which, once inherited by the child process memory space, often causes obscure bugs like segmentation fault. As a workaround, you can provide a dummy_env_fn that constructs an environment without initializing OpenGL.

The ReplayBuffer is a key data structure for storing a demonstration dataset both in-memory and on-disk with chunking and compression. It makes heavy use of the zarr format but also has a numpy backend for lower access overhead.

On disk, it can be stored as a nested directory (e.g. data/pusht_cchi_v7_replay.zarr) or a zip file (e.g. data/robomimic/datasets/square/mh/image_abs.hdf5.zarr.zip).

Due to the relative small size of our datasets, it's often possible to store the entire image-based dataset in RAM with Jpeg2000 compression which eliminates disk IO during training at the expense increasing of CPU workload.

Example:

data/pusht_cchi_v7_replay.zarr
 ├── data
 │   ├── action (25650, 2) float32
 │   ├── img (25650, 96, 96, 3) float32
 │   ├── keypoint (25650, 9, 2) float32
 │   ├── n_contacts (25650, 1) float32
 │   └── state (25650, 5) float32
 └── meta
     └── episode_ends (206,) int64

Each array in data stores one data field from all episodes concatenated along the first dimension (time). The meta/episode_ends array stores the end index for each episode along the fist dimension.

The SharedMemoryRingBuffer is a lock-free FILO data structure used extensively in our real robot implementation to utilize multiple CPU cores while avoiding pickle serialization and locking overhead for multiprocessing.Queue.

As an example, we would like to get the most recent To frames from 5 RealSense cameras. We launch 1 realsense SDK/pipeline per process using SingleRealsense, each continuously writes the captured images into a SharedMemoryRingBuffer shared with the main process. We can very quickly get the last To frames in the main process due to the FILO nature of SharedMemoryRingBuffer.

We also implemented SharedMemoryQueue for FIFO, which is used in RTDEInterpolationController.

In contrast to OpenAI Gym, our polices interact with the environment asynchronously. In RealEnv, the step method in gym is split into two methods: get_obs and exec_actions.

The get_obs method returns the latest observation from SharedMemoryRingBuffer as well as their corresponding timestamps. This method can be call at any time during an evaluation episode.

The exec_actions method accepts a sequence of actions and timestamps for the expected time of execution for each step. Once called, the actions are simply enqueued to the RTDEInterpolationController, and the method returns without blocking for execution.

Read and imitate:

• diffusion_policy/dataset/pusht_image_dataset.py
• diffusion_policy/env_runner/pusht_image_runner.py
• diffusion_policy/config/task/pusht_image.yaml

Make sure that shape_meta correspond to input and output shapes for your task. Make sure env_runner._target_ and dataset._target_ point to the new classes you have added. When training, add task=<your_task_name> to train.py's arguments.

Read and imitate:

• diffusion_policy/workspace/train_diffusion_unet_image_workspace.py
• diffusion_policy/policy/diffusion_unet_image_policy.py
• diffusion_policy/config/train_diffusion_unet_image_workspace.yaml

Make sure your workspace yaml's _target_ points to the new workspace class you created.

This repository is released under the MIT license. See LICENSE for additional details.

• Our ConditionalUnet1D implementation is adapted from Planning with Diffusion.
• Our TransformerForDiffusion implementation is adapted from MinGPT.
• The BET baseline is adapted from its original repo.
• The IBC baseline is adapted from Kevin Zakka's reimplementation.
• The Robomimic tasks and ObservationEncoder are used extensively in this project.
• The Push-T task is adapted from IBC.
• The Block Pushing task is adapted from BET and IBC.
• The Kitchen task is adapted from BET and Relay Policy Learning.
• Our shared_memory data structures are heavily inspired by shared-ndarray2.