FLD: Fourier Latent Dynamics for Structured Motion Representation and Learning

Authors: Chenhao Li, Elijah Stanger-Jones, Steve Heim, Sang bae Kim

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate FLD on the MIT Humanoid robot (Chignoli et al., 2021), with which we show its applicability to state-of-the-art real-world robotic systems. We use the human locomotion clips collected in Peng et al. (2018) retargeted to the joint space of our robot as the reference motion dataset... We compare the motion embeddings of the reference dataset obtained from training FLD following Sec. 4.2 with different models... We illustrate the latent embeddings acquired by these models in Fig. 4... We demonstrate the generality of FLD in reconstructing and predicting unseen motions during training. Figure 5 (left) illustrates a representative validation with a diagonal run motion... We perform an online tracking experiment... We perform ablation studies with different skill sampler implementations...
Researcher Affiliation Academia Chenhao Li, Elijah Stanger-Jones, Steve Heim, Sangbae Kim Department of Mechanical Engineering, Massachusetts Institute of Technology {chenhli, elijahsj, sheim, sangbae}@mit.edu
Pseudocode Yes Algorithm 1 Policy training
Open Source Code Yes Supplementary videos and more details for this work are available at https://sites.google.com/view/iclr2024-fld/home. ... The code has been open-sourced on the project page.
Open Datasets Yes We use the human locomotion clips collected in Peng et al. (2018) retargeted to the joint space of our robot as the reference motion dataset...
Dataset Splits No No explicit train/validation/test dataset splits (e.g., percentages or sample counts) were provided. The paper mentions training details like "max iterations", "learning epochs", "mini-batches", and "parallel training environments" but not how the dataset itself was partitioned.
Hardware Specification No The paper mentions using "Py Torch 1.10 with CUDA 12.0" and that "All training is done by collecting experiences from 4096 uncorrelated instances of the simulator in parallel." and acknowledges "MIT Super Cloud and Lincoln Laboratory Supercomputing Center for providing HPC resources." However, it does not provide specific hardware details such as GPU/CPU models, their quantities, or memory specifications.
Software Dependencies Yes The learning networks and algorithm are implemented in Py Torch 1.10 with CUDA 12.0. Adam is used as the optimizer for training the representation models.
Experiment Setup Yes The information is summarized in Table S3. ... Table S3: Representation training parameters Parameter Symbol Value step time seconds t 0.02 max iterations 5000 learning rate 0.0001 weight decay 0.0005 learning epochs 5 mini-batches 10 latent channels c 8 trajectory segment length H 51 FLD propagation horizon N 50 propagation decay α 1.0 approximate training hours 1 ... The information is summarized in Table S8. ... Table S8: Policy training parameters Parameter Symbol Value step time seconds t 0.02 skill-performance buffer size |B| 5000 max episode time seconds 20 max iterations 20000 learning rate 0.001 steps per iteration 24 learning epochs 5 mini-batches 4 KL divergence target 0.01 discount factor γ 0.99 clip range ϵ 0.2 entropy coefficient 0.01 parallel training environments 4096 number of seeds 5 approximate training hours 2