HITL-TAMP: Human-In-The-Loop Task and Motion Planning for Imitation Learning

Ajay Mandlekar*1    Caelan Garrett*1    Danfei Xu1, 2    Dieter Fox1
1NVIDIA       2Georgia Institute of Technology
Published at CoRL 2023.

HITL-TAMP combines the benefits of imitation learning and TAMP to solve contact-rich and long-horizon manipulation



During data collection, the TAMP system defers task portions it does not know how to complete to a human operator. This greatly reduces the data collection burden on the human operator, who does not need to demonstrate the full task.




During deployment, TAMP defers task portions to an autonomous robot trained via imitation learning on the collected data. This enables TAMP to work for contact-rich manipulation without accurate models.




Long-horizon tasks can be difficult for standard imitation learning approaches. By contrast, the HITL-TAMP framework easily scales to long-horizon tasks such as this coffee preparation task, which consists of 8 stages (4 TAMP and 4 policy segments).


HITL-TAMP trains proficient real-world agents on challenging contact-rich and long-horizon manipulation

HITL-TAMP policy can solve the challenging Tool Hang task.

HITL-TAMP policy is robust to different initializations.

Stack Three

Coffee Broad

Coffee

HITL-TAMP greatly accelerates data collection and often trains near-perfect agents from this data

Queueing System

Since humans only need to demonstrate small pieces of each task, they can participate in several data collection sessions at once. Human operators interact asynchronously with a fleet of several robots using our queueing system, which schedules tasks to ensure humans are always kept busy, enabling large-scale data collection.

HITL-TAMP Queueing System

Dataset Visualizations

We collected 2.1K+ demos using the system across these tasks. This took a single operator a few hours, as opposed to operator teams and days needed in prior work. Similarly, in our user study, given a fixed time budget, users collected much more data with HITL-TAMP (over 3x more) compared to conventional teleoperation.

Square

Square Broad

Three Piece Assembly

Three Piece Assembly Broad

Coffee

Coffee Broad

Coffee Preparation

Tool Hang

Tool Hang Broad

Trained Policies

Each policy below was trained on 200 demos from HITL-TAMP. Several policies are near-perfect.

Square (100%)

Square Broad (100%)

Three Piece Assembly (100%)

Three Piece Assembly Broad (85%)

Coffee (100%)

Coffee Broad (99%)

Coffee Preparation (96%)

Tool Hang (81%)

Tool Hang Broad (49%)

HITL-TAMP can train performant agents from just 10 minutes of data provided by operators with little to no teleoperation experience

The policies below were trained on just 10 minutes of operator data from an operator with little to no experience with teleoperation. We compare policies trained on 10 minutes of HITL-TAMP data and 10 minutes of conventional teleoperation data.

Coffee

HITL-TAMP (100%)

Conventional (28%)

Square Broad

HITL-TAMP (84%)

Conventional (0%)

Three Piece Assembly Broad

HITL-TAMP (22%)

Conventional (0%)

Task Reset Distributions

Below, we show the reset distributions for each task.

Square

Square Broad

Three Piece Assembly

Three Piece Assembly Broad

Coffee

Coffee Broad

Coffee Preparation

Tool Hang

Tool Hang Broad

Stack Three Real

Coffee Real

Coffee Broad Real

Tool Hang Real

Real World Data Collection

The videos below are from the HITL-TAMP datasets that we used to train our real world agents. The video resolutions match the image resolutions used for policy training.

Stack Three Real

Coffee Real

Coffee Broad Real

Tool Hang Real

Full Supplemental Video

BibTeX

@inproceedings{mandlekar2023hitltamp,
    title={Human-In-The-Loop Task and Motion Planning for Imitation Learning},
    author={Mandlekar, Ajay and Garrett, Caelan and Xu, Danfei and Fox, Dieter},
    booktitle={7th Annual Conference on Robot Learning},
    year={2023}
}

Acknowledgements

This work was made possible due to the help and support of Sandeep Desai (robot hardware), Ravinder Singh (IT), Alperen Degirmenci (compute cluster), Yashraj Narang (valuable discussions), Anima Anandkumar (access to robot hardware), Yifeng Zhu (robot control framework), and Shuo Cheng (drawer design used in Coffee Preparation task). We also thank all of the participants of our user study for contributing their valuable time and feedback on their experience.