Skip to content

quick_start.md

Latest commit

 

History

History
60 lines (51 loc) · 2.27 KB

quick_start.md

File metadata and controls

60 lines (51 loc) · 2.27 KB
Raw

Quick start

This quick start allows you to collect data in the MuJoCo simulation and train and rollout the ACT policy.

Install

Install RoboManipBaselines:

$ git clone git@github.com:isri-aist/RoboManipBaselines.git --recursive
$ cd RoboManipBaselines
$ pip install -e .[act]

Install ACT from a third party:

$ cd third_party/act/detr
$ pip install -e .

Data collection by teleoperation

Tip

Instead of collecting data by teleoperation, you can download the public dataset TeleopMujocoUR5eCable_Dataset30 from here.

Operate the robot in the simulation and save the data:

# Go to the top directory of this repository
$ cd robo_manip_baselines
$ # Connect a SpaceMouse to your PC
$ python ./bin/Teleop.py MujocoUR5eCable --world_idx_list 0 5 --input_device keyboard

Tip

A teleoperation input device such as a 3D mouse can be used instead of a keyboard. See here.

In our experience, models can be trained stably with roughly 30 data sets. The teleoperation data is saved in the robo_manip_baselines/dataset/MujocoUR5eCable_<date_suffix> directory (e.g., MujocoUR5eCable_20240101_120000).

Model training

Train the ACT:

# Go to the top directory of this repository
$ cd robo_manip_baselines
$ python ./bin/Train.py Act --dataset_dir ./dataset/MujocoUR5eCable_20240101_120000

The learned parameters are saved in the robo_manip_baselines/checkpoint/Act/<dataset_name>_Act_<date_suffix> directory (e.g., MujocoUR5eCable_20240101_120000_Act_20240101_130000).

Note

The following error will occur if the chunk_size is larger than the time series length of the training data. In such a case, either set the --skip option to a small value, or set the --chunk_size option to a small value.

RuntimeError: The size of tensor a (70) must match the size of tensor b (102) at non-singleton dimension 0

Policy rollout

Rollout the ACT in the simulation:

# Go to the top directory of this repository
$ cd robo_manip_baselines
$ python ./bin/Rollout.py Act MujocoUR5eCable \
--checkpoint ./checkpoint/Act/MujocoUR5eCable_20240101_120000_Act_20240101_130000/policy_last.ckpt \
--world_idx 0