This is the official implementation of the paper Watch-And-Help: A Challenge for Social Perception and Human-AI Collaboration.
In this work, we introduce Watch-And-Help (WAH), a challenge for testing social intelligence in agents. In WAH, an AI agent needs to help a human-like agent perform a complex household task efficiently. To succeed, the AI agent needs to i) understand the underlying goal of the task by watching a single demonstration of the human-like agent performing the same task (social perception), and ii) coordinate with the human-like agent to solve the task in an unseen environment as fast as possible (human-AI collaboration).
We provide a dataset of tasks to evaluate the challenge, as well as different baselines consisting on learning and planning-based agents.
Check out a video of the work here.
If you use this code in your research, please consider citing.
@inproceedings{
puig2021watchandhelp,
title={Watch-And-Help: A Challenge for Social Perception and Human-{\{}AI{\}} Collaboration},
author={Xavier Puig and Tianmin Shu and Shuang Li and Zilin Wang and Yuan-Hong Liao and Joshua B. Tenenbaum and Sanja Fidler and Antonio Torralba},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=w_7JMpGZRh0}
}
Clone the VirtualHome API repository one folder above this repository
cd ..
git clone --branch wah https://github.com/xavierpuigf/virtualhome.git
cd virtualhome
pip install -r requirements.txtDownload the simulator, and put it in an executable folder, one folder above this repository
[IMPORTANT] Please make sure to use the wah branch of the VirtualHome repo and the v2.2.0 version of the executable.
pip install -r requirements.txtDownload the checkpoints for the weights to the hybrid help model and the watching model. Use
sh scripts/download_ckpts.shWe include a dataset of environments and activities that agents have to perform in them. During the Watch phase and the training of the Help phase, we use a dataset of 5 environments. When evaluating the Help phase, we use a dataset of 2 held out environments.
The Watch phase consists of a set of episodes in 5 environments showing Alice performing the task. These episodes were generated using a planner, and they can be downloaded here. The training and testing split information can be found in datasets/watch_scenes_split.json.
The Help phase, contains a set of environments and tasks definitions. You can find the train and test datasets used in dataset/train_env_set_help.pik and dataset/test_env_set_help.pik. Note that the train environments are independent, whereas the test environments match the tasks in the Watch test split.
You can also create your dataset, and modify it to incorporate new tasks. For that, run
python gen_data/vh_init.py --num-per-apartment {NUM_APT} --task {TASK_NAME}Where NUM_APT corresponds to the number of episodes you want for each apartment and task and TASK_NAME corresponds to the task name you want to generate, which can be setup_table, clean_table, put_fridge, prepare_food, read_book, watch_tv or all to generate all the tasks.
After creating your dataset, you can create the data for the Watch phase running the Alice alone baseline (see Evaluate Baselines).
You can then generate a dataset of tasks in a new environment where the tasks match those of the Watch phase. We do that in our work to make sure that the environment in the Watch phase is different than that in the Help Phase while having the same task specification. You can do that by running:
python gen_data/vh_init_gen_test.pyIt will use the tasks from the test split of the Watch phase to create a Help dataset.
First, download the dataset for the Watch phase and put it under dataset.
You can train the goal prediction model for the Watch phase as follows:
sh scripts/train_watch.shTo test the goal prediction model, run:
sh scripts/test_watch.shWe provide planning and learning-based agents for the Helping stage. The agents have partial observability in the environment, and plan according to a belief that updates with new observations.
We train the hybrid baseline in 2 stages. One where we allow the agent to teleport to a location and one where the agent has to walk to the location. You can train the hybrid baseline as follows.
# First stage
python training_agents/train_a2c.py \
--max-num-edges 10 --max-episode-length 250 \
--batch_size 32 --obs_type partial --gamma 0.95 \
--lr 1e-4 --nb_episodes 100000 --save-interval 200
--simulator-type unity --base_net TF --log-interval 1 \
--long-log 50 --logging --base-port 8681
--num-processes 5 --teleport
--executable_file ../executable/linux_exec_v3.x86_64 \
--agent_type hrl_mcts --num_steps_mcts 24
#Second stage
python training_agents/train_a2c.py \
--max-num-edges 10 --max-episode-length 250 \
--batch_size 32 --obs_type partial --gamma 0.95 \
--lr 1e-4 --nb_episodes 100000 --save-interval 200 \
--simulator-type unity --base_net TF --log-interval 1 \
--long-log 50 --logging --base-port 8681 \
--num-processes 5 \
--executable_file ../executable/linux_exec_v3.x86_64 \
--agent_type hrl_mcts --num_steps_mcts 50 \
--load-model {path to previous model}
Below is the code to evaluate the different planning-based models. The results will be saved in a folder called test_results. Make sure you create it first, one level above this repository.
mkdir ../test_resultsHere is the code to evaluate the baselines proposed in the paper.
# Alice alone
python testing_agents/test_single_agent.py
# Bob planner true goal
python testing_agents/test_hp.py
# Bob planner predicted goal
python testing_agents/test_hp_pred_goal.py
# Bob planner random goal
python testing_agents/test_hp_random_goal.py
# Bob random actions
python testing_agents/test_random_action.pyBelow is the code to evaluate the learning-based methods
# Hybrid Baseline
CUDA_VISIBLE_DEVICES=0 python testing_agents/test_hybrid.py \
--max-num-edges 10 --max-episode-length 250 \
--num-processes 1 \
--agent_type hrl_mcts --num_steps_mcts 40 \
--load-model checkpoints/checkpoint_hybrid.pt
[Coming Soon] You may want to generate a video to visualize the episode you just generated. Here we include a script to view the episodes you generate during the Help phase.