This repo contains supplementary code for our paper Getting to Know One Another: Calibrating Intent, Capabilities and Trust for Human-Robot Collaboration (IROS, 2020).
- Setup
- Instructions for Basic Execution
- Instructions for Reproducing Simulation Experiment Results
- Experimental Setups
- Developer Guide
- Support
You will need to install PyPy3.
To execute TICC-MCP solver, run
pypy3 driver.py <seed>
To execute POMCP solver without capability models, run
pypy3 'standard POMCP/driver.py' <seed>
To reproduce experimental results for TICC-MCP solver, run
bash job_launcher.sh
To reproduce experimental results for POMCP solver without capability models, run
bash std_job_launcher.sh
Redirect sys.stdout to write output to log files.
The experimental setups can be adjusted in driver.py.
In this setup, the number of shopping lists is fixed at 10 and number of shopping item types is fixed at 5. The number of search samples is varied from 5000 to 50000. Vary num_iter in driver.py based on number of search samples.
The shopping lists and the actual capabilities setup are as follows:
Shopping lists
| Item 1 | Item 2 | Item 3 | Item 4 | Item 5 |
|---|---|---|---|---|
| 4 | 3 | 0 | 2 | 3 |
| 1 | 4 | 0 | 7 | 1 |
| 2 | 3 | 2 | 3 | 3 |
| 5 | 4 | 2 | 0 | 2 |
| 0 | 3 | 3 | 4 | 3 |
| 3 | 3 | 0 | 3 | 3 |
| 6 | 3 | 0 | 1 | 2 |
| 2 | 3 | 4 | 1 | 2 |
| 1 | 1 | 2 | 4 | 4 |
| 0 | 3 | 2 | 5 | 2 |
Human's capability
| Item 1 | Item 2 | Item 3 | Item 4 | Item 5 |
|---|---|---|---|---|
| 0% | 100% | 10% | 0% | 100% |
Robot's capability
| Item 1 | Item 2 | Item 3 | Item 4 | Item 5 |
|---|---|---|---|---|
| 100% | 0% | 100% | 100% | 10% |
In this setup, the number of shopping lists is fixed at 2 and number of search samples is fixed at 50000. The number of shopping item types is varied from 2 to 5. Vary reward_space in driver.py accordingly.
The shopping lists and the actual capabilities setup are as follows:
Shopping lists
| Item 1 | Item 2 |
|---|---|
| 1 | 12 |
| 2 | 10 |
Human's capability
| Item 1 | Item 2 |
|---|---|
| 50% | 100% |
Robot's capability
| Item 1 | Item 2 |
|---|---|
| 100% | 50% |
Shopping lists
| Item 1 | Item 2 | Item 3 |
|---|---|---|
| 8 | 5 | 0 |
| 2 | 5 | 6 |
Human's capability
| Item 1 | Item 2 | Item 3 |
|---|---|---|
| 0% | 100% | 10% |
Robot's capability
| Item 1 | Item 2 | Item 3 |
|---|---|---|
| 100% | 0% | 100% |
Shopping lists
| Item 1 | Item 2 | Item 3 | Item 4 |
|---|---|---|---|
| 4 | 4 | 2 | 3 |
| 3 | 5 | 0 | 5 |
Human's capability
| Item 1 | Item 2 | Item 3 | Item 4 |
|---|---|---|---|
| 0% | 100% | 10% | 100% |
Robot's capability
| Item 1 | Item 2 | Item 3 | Item 4 |
|---|---|---|---|
| 100% | 0% | 100% | 10% |
Shopping lists
| Item 1 | Item 2 | Item 3 | Item 4 | Item 5 |
|---|---|---|---|---|
| 2 | 3 | 2 | 3 | 3 |
| 5 | 4 | 2 | 0 | 2 |
Human's capability
| Item 1 | Item 2 | Item 3 | Item 4 | Item 5 |
|---|---|---|---|---|
| 0% | 100% | 10% | 0% | 100% |
Robot's capability
| Item 1 | Item 2 | Item 3 | Item 4 | Item 5 |
|---|---|---|---|---|
| 100% | 0% | 100% | 100% | 10% |
In this setup, the number of shopping items is fixed at 5 and number of search samples is fixed at 50000. The number of shopping lists is varied from 5 to 10.
For an n-shopping list setup, use the first n shopping lists used for Setup 1. Vary reward_space in driver.py accordingly. The actual capabilities setup is also the same as Setup 1.
The main solver program is located at the root level. It is modularized into several main components:
- Driver: The driver coordinates the functions of the other components. It first takes in input parameters and sets up the TICC-POMDP environment and the TICC-MCP solver. It then executes the solver. After every robot action, the driver takes in a human action input from either an actual human or a simulated human. It is also responsible in returning the intermediate and final results.
- Solver: The solver implements the TICC-MCP algorithm. It is responsible for executing the search, as well as building and maintaining the search tree.
- Search Tree: The search tree comprises two types of nodes: 1) Robot action node which represents an action taken by the robot 2) Human action node which represents an action taken by the human. It keeps track of the necessary statistics and stores the tree structure.
- Environment: The environment contains the TICC-POMDP model in which the robot and human operate. It has all the essential physics and information of the world.
The high level architecture is as follows:
A similar implementation is done for the standard POMCP algorithm without capability models, which can be found under /standard POMCP. To ensure the correctness of the program, there are unit tests for all testable atomic functions, which can be found under /tests.
For better efficiency, the sparse capability model representation can be reduced a flatten list of important entries (two entries for each action).
For further questions, reach out to me at joshualeekaisheng@gmail.com.
If you use this code or find it useful, please cite our paper:
@inproceedings{lee2020trust,
title={Getting to Know One Another: Calibrating Intent, Capabilities and Trust for Human-Robot Collaboration},
author={Joshua Lee and Jeffrey Fong and Bing Cai Kok and and Harold Soh},
booktitle = {{IEEE/RSJ International Conference on Intelligent Robots and Systems} (Accepted)},
year = {2020},
month = {October}}