BAKU: An Efficient Transformer for Multi-Task Policy Learning
Authors: Siddhant Haldar, Zhuoran Peng, Lerrel Pinto
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on 129 simulated tasks across LIBERO, Meta-World suite, and the Deepmind Control suite exhibit an overall 18% absolute improvement over RT-1 and MT-ACT, with a 36% improvement on the harder LIBERO benchmark. On 30 real-world manipulation tasks, given an average of just 17 demonstrations per task, BAKU achieves a 91% success rate. |
| Researcher Affiliation | Academia | Siddhant Haldar Zhuoran Peng Lerrel Pinto New York University |
| Pseudocode | No | The paper describes architectural components and algorithmic ideas but does not contain a structured pseudocode or algorithm block. |
| Open Source Code | Yes | All of our datasets, and training and evaluation code will be made publicly available. Videos of our trained policies can be seen here: baku-robot.github.io. |
| Open Datasets | Yes | We experiment with 90 manipulation tasks from the LIBERO-90 benchmark [34], 30 manipulation tasks from Meta-World suite [76], and 9 locomotion tasks from Deep Mind Control Suite (DMC) [67]. |
| Dataset Splits | No | The paper discusses training and test phases but does not explicitly provide details about a validation dataset split, its size, or how it was used. |
| Hardware Specification | Yes | Training time Below we provide details about the time required to train BAKU on a single NVIDIA RTX A4000 GPU. |
| Software Dependencies | Yes | Transformer architecture min GPT [29] (with 8 layers and 4 heads) |
| Experiment Setup | Yes | The complete list of hyperparameters is provided in Table 4. For RT-1 [6], we use our implementation with an RT-1 action head that discretizes the continuous action into discrete bins uniformly. For MT-ACT [5], we use the open-source implementation with the default hyperparameters. We vary the action chunk length for MT-ACT for different benchmarks, the values for which have been provided in Table 4. |