Motion Control of High-Dimensional Musculoskeletal Systems with Hierarchical Model-Based Planning
Authors: Yunyue Wei, Shanning Zhuang, Vincent Zhuang, Yanan Sui
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we aim to comprehensively evaluate MPC2, and seek to answer the following questions: 1) Can MPC2 achieve robust and performant control over a wide variety of motion tasks and models? ... We implement MPC2 using the Mujoco MPC (MJPC) platform... We show control sequences of the Stand, Walk, Rough, Slope, and Stair tasks using MPC2 in Figure 3. ... In the Stand and Walk tasks, we compared the control performance of MPC2 with the current state-of-the-art DRL-based algorithms... Ablation studies over the algorithm components further verify the principled design and biomechanical fidelity of MPC2. |
| Researcher Affiliation | Collaboration | Yunyue Wei1, Shanning Zhuang1, Vincent Zhuang2, Yanan Sui1 1 Tsinghua University 2 Google Deep Mind EMAIL EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1: Model Predictive Control with Morphology-aware Proportional Control (MPC2) Input: Model dynamics f, rollout horizon H, total rollout number N, instant rollout number N, iteration number r, distribution parameter µ, σ, current state st 1 for i = 1, , r do 2 z1, ..., z N N(Mpos(st), σ) // Instant rollout 3 z N+1, ..., z N N(µ, σ) // MPPI rollout 4 C1 θ, , CN θ RMP(z1, H), , RMP(z N, H) 5 Update µ, σ using e.q. (6) 7 z µ, ˆuθ t πMP(st, z ) 8 return ˆuθ t , z , µ, σ |
| Open Source Code | No | The paper does not explicitly state that the source code for the methodology described is provided, nor does it include a link to a code repository for MPC2. It mentions a 'project page' for videos and uses external open-source tools like MyoSuite and Mujoco MPC, but not their own code. |
| Open Datasets | Yes | Our primary experiments are conducted on the MS-Human-700, a comprehensive wholebody musculoskeletal model comprising 90 rigid body segments, 206 joints, and 700 muscle-tendon units (Zuo et al., 2024). Additionally, we employ an upper limb model of the human body and an ostrich model (La Barbera et al., 2021) to showcase the generalization of MPC2 across different models and tasks. |
| Dataset Splits | No | The paper describes experiments conducted on musculoskeletal simulation models (MS-Human-700, an ostrich model, and an arm model) within the MuJoCo physics simulator. It does not involve traditional empirical datasets that would require training, validation, and test splits for machine learning tasks. |
| Hardware Specification | Yes | The experiments of MPC2 were conducted on a server equipped with an AMD EPYC 7773X processor, an NVIDIA Ge Force RTX 4090 GPU, and 512 GB of memory. |
| Software Dependencies | No | The paper mentions using the 'Mujoco MPC (MJPC) platform' and for baselines 'Stable baselines3' but does not provide specific version numbers for these software dependencies or any other key libraries. |
| Experiment Setup | Yes | In all experiments, we set the iteration number r of MPC2 to 1 for rapid response to the changing states in the main thread, and sample 64 rollouts (containing N = 10 instant rollouts) across a 0.3s horizon during each round of planning. Unless otherwise noted, the simulation in main thread are run with 20% of the real-time speed... The cost function design of all movement tasks is detailed in Appendix B. Appendix B details specific cost function parameters like 'Htarget = 1.55 vtarget = 0' for Stand task. |