Awesome

Chain-of-Thought Predictive Control (CoTPC)

This is the official repository for CoTPC, a powerful hierarchical imitation learning model presented in the following paper:

Chain-of-Thought Predictive Control<br>

Zhiwei Jia, Fangchen Liu, Vineet Thumuluri, Linghao Chen, Zhiao Huang, Hao Su<br> UC San Diego, UC Berkeley, Zhejiang University<br>

Tasks

Currently the code supports five tasks from the ManiSkill2 benchmark: PickCube-v0, StackCube-v0, PegInsertionSide-v0, TurnFaucet-v0, and PushChair-v1.

Demonstration Data

The state-based demo trajectories used in the paper are stored in this Googld Drive folder. Each folder has a *.h5 file (for actual trajectories) and a *.json file (for metadata regarding the trajectories). Each task has over 1000 demo trajectories. Each trajectory comes with a different env configurations (i.e., env seed, which influences object poses, object geometries, etc.). These demos are generated by replaying the official ManiSkill2 demos or the ones adapted from this repo with several patches to the ManiSkill2 code (see CoTPC/maniskill2_patches). Specifically, we add additional flags to the tasks so that the key states (the Chain-of-Thought) can be obtained with priviledged information from the simulator. For the task TurnFaucet-v0, we use a subset of 10 faucet models for the demos (see CoTPC/scripts/replay_turn_faucet_trajectories.sh). For the task PushChair-v1, we use a subset of 5 chair models for the demos. If you want to generate visual-based demos, please refer to the official ManiSkill2 guidance here.

Data Loader

The data loader samples (contiguous) subarrays of demo trajectories by specifying the min and max sample lengths. In CoTPC, we simply use a fixed value for both min and max (e.g., set the context size as 60 for all of the 5 tasks from ManiSkill2). With a fixed random seed seed and num_traj, each time the data loader samples a fixed subset of all trajectories (by default, we use seed=0). For TurnFaucet and PushChair, due to the variations of the different object models, the loader performs sampling such that the number of trajs per model is the same (hopefully this data balance eases model training). The key states (a.k.a the Chain-of-Thought) can be obtained with the function get_key_states(data_idx) that accesses privileged info available during training.

Evaluation

We patch the environment code from ManiSkill2 (see CoTPC/maniskill2_patches) to provide additional evaluation metrics (intermediate success rate) each task execution. We report evaluation results for models using both the seen env configurations and the unseen env configuration. For tasks involving geometric variations we also evaluate using unseen objects (i.e., zero-shot transfer). Please refer to CoTPC/src/eval.py for details.

Sampled Training Scripts

Please see CoTPC/scripts/train.sh as examples.