MeMo: Meaningful, Modular Controllers via Noise Injection
Authors: Megan Tjandrasuwita, Jie Xu, Armando Solar-Lezama, Wojciech Matusik
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We benchmark our framework in locomotion and grasping environments on simple to complex robot morphology transfer. We also show that the modules help in task transfer. On both structure and task transfer, Me Mo achieves improved training efficiency to graph neural network and Transformer baselines. |
| Researcher Affiliation | Collaboration | Megan Tjandrasuwita MIT megantj@mit.edu Jie Xu NVIDIA jiex@nvidia.com Armando Solar-Lezama MIT asolar@csail.mit.edu Wojciech Matusik MIT wojciech@mit.edu |
| Pseudocode | No | The paper describes the approach and training pipeline but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code can be found at https://github.com/MeganTj/MeMo. |
| Open Datasets | No | The paper refers to environments generated by the Robo Grammar [16] and Diff Red Max [17] simulators, but does not provide explicit links or citations for a pre-existing dataset used for training. These are frameworks used to construct tasks. |
| Dataset Splits | Yes | We sample 500 trajectories from the expert controllers of the 6 leg centipede, 6 leg worm, and 6 leg hybrid and 250 trajectories from the controller of the 4 finger claw as the validation sets. |
| Hardware Specification | Yes | We run experiments on 2 different machines with AMD Ryzen Threadripper PRO 3995WX processors and NVIDIA RTX A6000 GPUs. Both machines have 64 CPU cores and 128 threads. |
| Software Dependencies | No | The paper mentions using PyTorch for implementations and specific codebases for PPO and Nerve Net, but does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | We conduct an extensive hyperparameter search and find that that the values in Table 2 yield reasonable performance. For imitation learning, we use a batch size of 1024 and tune the learning rate in [7e-4, 1e-3, 2e-3, 4e-3, 7e-3]. |