DittoGym: Learning to Control Soft Shape-Shifting Robots
Authors: Suning Huang, Boyuan Chen, Huazhe Xu, Vincent Sitzmann
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we hope to answer the following questions via empirical evaluation. 1) Can our formalization, when combined with appropriate algorithms, enable simulated soft robots to perform interesting tasks that require fine-grained morphology changes? 2) Can our benchmark Ditto Gym sufficiently evaluate algorithms designed to control fine-grained morphology changes for soft robots? 3) Compared to relevant baselines, how sample efficient is CFP for reconfigurable soft robot control? In Figure 6, we plot the episode reward curves for all the runs. Our proposed algorithm, CFP, outperforms all baselines consistently across all tasks in terms of overall sample efficiency. In addition, CFP can consistently attain higher episode reward upon convergence. |
| Researcher Affiliation | Academia | Suning Huang Department of Automation Tsinghua University hsn19@mails.tsinghua.edu.cn Boyuan Chen CSAIL Massachusetts Institute of Technology boyuanc@mit.edu Huazhe Xu IIIS Tsinghua University huazhe_xu@mail.tsinghua.edu.cn Vincent Sitzmann CSAIL Massachusetts Institute of Technology sitzmann@mit.edu |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at Ditto Gym and CFP. More results are available at https://dittogym.github.io. |
| Open Datasets | Yes | We introduce Ditto Gym, a comprehensive RL benchmark for reconfigurable soft robots that require fine-grained morphology changes to accomplish the tasks. Our benchmark, Ditto Gym, addresses the absence of standardized benchmarks for reconfigurable soft robots with fine-grained morphology changes. We incorporate our formalized simulation (Section 4.1) into a set of Open AI Gym environments. |
| Dataset Splits | No | The paper describes training and evaluation within a reinforcement learning benchmark, but does not provide specific training/validation/test dataset splits in terms of percentages or sample counts for static data. |
| Hardware Specification | Yes | If simulation alone, the environment can run at an average speed of 70.1 FPS on the RTX4090 GPU, with only 20-30% volatile utility usage and 15% memory usage. Training a single agent runs at approximately 20 FPS when executed on the same GPU. |
| Software Dependencies | No | The paper mentions software like SAC and Taichi, and refers to frameworks like Raffin et al., 2019 and Huang et al., 2021, but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | The specific parameters of the reinforcement learning algorithm we employed are presented in Table 3. Batch Size 256 Auto Entropy-tuning False Gamma 0.99 Critic Network Hidden Size 256 Learning Rate 3e-4 Momentum 0.99 Optimizer Adam Replay Buffer Size 200000 Seed 0,1 and 2. |