Chain-of-Thought Predictive Control
Authors: Zhiwei Jia, Vineet Thumuluri, Fangchen Liu, Linghao Chen, Zhiao Huang, Hao Su
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Co TPC on several challenging low-level control tasks (Moving Maze, Franka-Kitch and Mani Skill2) and verify its design with ablation studies. We find that Co TPC consistently outperforms several strong baselines. |
| Researcher Affiliation | Academia | 1UC San Diego 2UC Berkeley 3Zhejiang University. Correspondence to: Zhiwei Jia <zjia@ucsd.edu>. |
| Pseudocode | No | The paper includes figures illustrating the architecture and mathematical formulations, but no explicit pseudocode blocks or algorithms labeled as such. |
| Open Source Code | No | The paper mentions "See project page" in the abstract and appendix, but does not provide a direct link to a code repository or an explicit statement confirming the release of the source code for the methodology. |
| Open Datasets | Yes | For Mani Skill2, we use the original demonstrations provided by Mani Skill2, which are generated by a mixture of TAMP solvers and heuristics. Please see the original paper for details (the actual code used to generate these demonstrations is not released, though). ... We replay a subset of the human demonstrations originally proposed in (Gupta et al., 2019) in the simulator. |
| Dataset Splits | Yes | We train all methods on 400 demo trajectories of different env. configs and evaluate on 100 unseen ones (results in Tab. 1). ... For the main results of all methods on Mani Skill2, we report the best performance among 3 training runs over the last 20 checkpoints, an eval protocol adopted by (Chi et al., 2023). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory specifications) used for running the experiments. |
| Software Dependencies | No | The paper describes various models and their configurations but does not provide specific version numbers for software dependencies (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We train it with a constant learning rate of 1e-3 with Adam optimizer with a batch size of 32 for 150K iterations (Moving Maze), 300K iterations (Franka Kitchen) and 500K iterations (Mani Skill2). ... The Transformer backbone has approximately the same number of parameters (~1M) as Co TPC and DT. We train the model for around 50k iterations... We use a coefficient of 100 for training the action offset regressor... |