metadata

task_categories:
  - robotics
tags:
  - LeRobot
  - RoboGrasp2024
  - Virtual
pretty_name: RoboGrasp 2024 Hackathon Dataset

2024 RoboGrasp Dataset

This is dataset the dataset for the RoboGrasp Hackathon 2024. It includes 108 simulated pick-and-place robot episodes collected on a simulated mobile aloha environment through teleoperated demonstrations. During the collected episodes the right arm of the robot is used to pick an item from the table and put it in a box on the top of the table.

There are three types of items:

green cube (47% of episodes)
red sphere (30% of episodes)
blue cylinder (22% of episodes)

You can visualize the dataset episodes here

Hackathon Task

Clone Repository

As a first step for the hackathon: clone the hackathon repository on the hackathon branch on your system. It's a fork of the original LeRobot repository containing the assets necessary for running following tasks.

git clone -b hackathon https://github.com/HumanoidTeam/lerobot-hackathon.git

Install Dependencies

After cloning the repository you can proceed in installing the dependencies using poetry. you can install poetry by running pip install poetry.

cd lerobot-hackathon
poetry lock
poetry build
poetry install --extras "gym-aloha"

This will create the virtual environment with all the dependencies required. You can source the environment by running poetry shell within the folder. From this point, you can configure and run your policy training using all the models present in Lerobot (e.g. ACT, DiffusionPolicy, VQ-Bet, etc...).

Policy Configuration

You can create a yaml file within the folder lerobot/configs/policy/. For example robograsp2024_submission_model.yaml. Within the the yaml file you can configure the input/output data shapes, data normalization strategies, context length and policy parameters. Some parts of the yaml file will be dependant on this dataset. Thus, we provide the parameters necessary to use this dataset.

Here are working configurations for the input and output structure: These go at the beginning of the yaml:

seed: 100000
dataset_repo_id: HumanoidTeam/robograsp_hackathon_2024

override_dataset_stats:
  observation.images.left_wrist:
    # stats from imagenet, since we use a pretrained vision model
    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
  observation.images.right_wrist:
    # stats from imagenet, since we use a pretrained vision model
    mean: [ [ [ 0.485 ] ], [ [ 0.456 ] ], [ [ 0.406 ] ] ]  # (c,1,1)
    std: [ [ [ 0.229 ] ], [ [ 0.224 ] ], [ [ 0.225 ] ] ]  # (c,1,1)
  observation.images.top:
    # stats from imagenet, since we use a pretrained vision model
    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)

These go within the policy: scope and regards the input/output datashape

  input_shapes:
    observation.images.left_wrist: [3, 480, 640]
    observation.images.right_wrist: [3, 480, 640]
    observation.images.top: [3, 480, 640]
    observation.state: ["${env.state_dim}"]
  output_shapes:
    action: ["${env.action_dim}"]

  # Normalization / Unnormalization
  input_normalization_modes:
    observation.images.left_wrist: mean_std
    observation.images.right_wrist: mean_std
    observation.images.top: mean_std
    observation.state: min_max
  output_normalization_modes:
    action: min_max

The remaining configuration can be derived from the other examples provided by the lerobot original repo.

Start Policy Training

you can start the policy training by running the following command, while having sourced the environment built in the previous section. To source the environment run:

poetry shell

To start the training, you can use this command:

 MUJOCO_GL="egl" python lerobot/scripts/train.py \                
    policy=robograsp2024_submission_model \
    env=humanoid_hackathon_mobile_aloha \
    env.task=AlohaHackathon-v0 \
    dataset_repo_id=HumanoidTeam/robograsp_hackathon_2024

Where robograsp2024_submission_model is the name of the yaml file with the policy configuration, humanoid_hackathon_mobile_aloha is the provided yaml configuration for the mujoco environment to test the trained policies.

Resume policy training from a checkpoint.

Terminated training too early? No worries! you can resume training from a previous checkpoint by running:

 MUJOCO_GL="egl" python lerobot/scripts/train.py \                
    policy=robograsp2024_submission_model \
    env=humanoid_hackathon_mobile_aloha \
    env.task=AlohaHackathon-v0 \
    dataset_repo_id=HumanoidTeam/robograsp_hackathon_2024 \
    hydra.run.dir=OUTPUT_PATH \
    resume=true

Where OUTPUT_PATH is the path to the checkpoint folder. It should look something like outputs/train/2024-10-23/18-38-31_aloha_MODELTYPE_default

Upload trained policy checkpoint

After training the model you can upload it to Huggingface with:

huggingface-cli upload $hf_username/$repo_name PATH_TO_CHECKPOINT

where PATH_TO_CHECKPOINT is the folder containing the checkpoints of your training. it should look like outputs/train/2024-10-23/23-02-55_aloha_diffusion_default/checkpoints/015000.

Policy Evaluation

python hackathon/evaluate_pretrained_policy_hackathon.py --device cuda --pretrained-policy-name-or-path HumanoidTeam/hackathon_sim_aloha --num-videos 5  --num-rollouts 10

This dataset was created using 🤗 LeRobot.