Research map / 3D vision / Human pose estimation
ImDy: Human Inverse Dynamics from Imitated Observations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ImDy, a data-driven human inverse dynamics method that builds a large-scale torque dataset by retraining the PHC imitation controller to output actuation torques from AMASS motions, then trains a dynamics estimator on it and fine-tunes on AddBiomechanics.
Problem
Human inverse dynamics (reproducing joint torques from kinematic observations) is limited in scalability because conventional optimization-based ID needs expensive laboratory setups such as force plates and motion capture.
Approach
Interpret imitation learning controllers as an inverse-dynamics knowledge base: retrain a physics-based imitator (PHC) to track motions with accessible actuation torques (set_dof_actuation_force_tensor) so that the imitated observations yield large-scale torque labels for training a data-driven ID model.
Key results
- On the AddBiomechanics test set, the FD model's marker RMSE is 0.0194 for ImDyS versus 0.0302 for the baseline trained on AddBiomechanics only, and 0.0186 for Nimble.
- Zero-shot ImDyS predictions on AddBiomechanics knee torques show trends similar to ground truth but with magnitudes that differ substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pre-train on ImDy before AddBiomechanics training | FD model · AddBiomechanics test set | marker RMSE: 0.0302 -> 0.0194 | Helped |
Surprising findings
- Zero-shot ImDyS knee torque predictions follow the trend of ground truth but differ strongly in magnitude, with ImDy tending to be over-powerful compared to real humans.
Limitations
- ImDy data shows more severe jittering than the original PHC outputs.
- Failures on ankle and knee torques attributed to data distribution rarely covering the sample motions in gait-dominated AddBiomechanics.
Experimental setup
Datasets: AMASS AddBiomechanics KIT BABEL
Benchmarks: AddBiomechanics
Metrics: RMSE
Baselines: PHC Nimble AddBiomechanics baseline
Builds on: PHC AddBiomechanics AMASS
Closest papers by meaning
- Homogeneous Dynamics Space for Heterogeneous Humans (CVPR 2025 · similarity 0.87)
- Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos (NeurIPS 2024 · similarity 0.87)
- PhysPT: Physics-aware Pretrained Transformer for Estimating Human Dynamics from Monocular Videos (CVPR 2024 · similarity 0.86)
- InterMimic: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions (CVPR 2025 · similarity 0.85)
- DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion (CVPR 2024 · similarity 0.85)
- Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.