Latent Dynamics · 2026-11-03
Why Koopman Structure Does Not Help Differentiable World Models for Humanoid Whole-Body Tracking
Haizhou Ge, Mengshuo Wang, Yufei Jia, Lei Han, Lu Shi, Guyue Zhou, Ruqi Huang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors test whether Koopman structure improves a differentiable world model used to train a simulated Unitree G1 whole-body tracking policy via BPTT. They find no gain: return is at best equal (0.823 vs 0.830), Koopman has 1.3–1.6× higher one-step error, and training is ~15% slower. Lifted rollouts diverge unless re-lifting at every imagined step; then Koopman effectively becomes action-affine, which is not better than an MLP.
Why it matters
Differentiable world models rely on rollout structure because gradients flow through imagined dynamics. The paper provides a controlled comparison showing that Koopman’s defining constraints (fixed lift + linear evolution with action entering linearly/bilinearly) do not help for humanoid whole-body tracking, and that the effective action-affine restriction is the main source of accuracy loss.
Method
- Replace SuperMimic’s MLP world model with a deep bilinear Koopman world model, keeping the rest of the BPTT tracking pipeline unchanged; compare offline prediction on identical data and online tracking in simulation on Unitree G1.
- Test two rollout regimes: lifting once (using lifted linear dynamics across steps) vs re-lifting (re-encoding at every imagined step) with additional variants (e.g., bounded Koopman, different lifts).
- Analyze structural error by fitting one-step increments to an action-affine form around the behavior policy’s noise-free action (operating point), and test repairs that alter what the lift reads.
Limitation
These conclusions concern one simulated robot and the Koopman structure where the action enters only linearly or bilinearly with a fixed lift; models whose action enters through a nonlinear encoder, or whose observables depend on the input, are outside the paper’s scope.
Abstract (from arXiv)
We test whether Koopman structure helps a differentiable world model used to train a simulated Unitree G1 whole-body tracking policy by backpropagation through imagined rollouts, replacing the MLP world model of an otherwise unchanged pipeline. It does not: return is at best equal (0.823 versus 0.830), one-step error on identical data is 1.3 to 1.6 times the MLP's, and training is about 15% slower. Lifted rollouts diverge or fail to train policies; of the schemes tested, re-lifting at every imagined step trains best. Re-lifting never uses the lifted linear dynamics and leaves a model affine in the action, with state-dependent offset and slope; an action-affine MLP also predicts worse, so the accuracy cost belongs to this class. At a fixed state, the dynamics' curvature in the action explains only part of it. Across states, the policy's actions span a curved region, so an affine model fits policy data only as the tangent at the policy's action, the operating point: on G1, the MLP's tangent there is as accurate as the MLP; at the mean action, 31 times worse. The offset must encode, from the state alone, where the policy acts, whereas an MLP reads the action. Supplying the policy's action removes most of the remaining gap on G1, Ant and Humanoid. The only tested online repair approaching the MLP feeds the action into the nonlinear lift, abandoning the structure. For this use, the MLP world model remains the better choice.
Related papers
- SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model
- GreenWAM: A World Foundation Model for Robotics
- BeJEPA: A Goal-Conditioned Humanoid Behavior World Model via Unsupervised RL
- DiffWBC: Unlocking Steerable General Humanoid Motion Tracking via One-Step Diffusion Whole-Body Control
- $\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-03 · Indexed by Pub-AI: 2026-11-03 UTC
Why Koopman Structure Does Not Help Differentiable World Models for Humanoid Whole-Body Tracking
First announced by arXiv: 2026-11-03 · Original record ↗
- arxiv
- 2611.00951
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-03 · First detected: 2026-11-03 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.00951v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
We test whether Koopman structure helps a differentiable world model used to train a simulated Unitree G1 whole-body tracking policy by backpropagation through imagined rollouts, replacing the MLP world model of an otherwise unchanged pipeline. It does not: return is at best equal (0.823 versus 0.830), one-step error on identical data is 1.3 to 1.6 times the MLP's, and training is about 15% slower. Lifted rollouts diverge or fail to train policies; of the schemes tested, re-lifting at every imagined step trains best. Re-lifting never uses the lifted linear dynamics and leaves a model affine in the action, with state-dependent offset and slope; an action-affine MLP also predicts worse, so the accuracy cost belongs to this class. At a fixed state, the dynamics' curvature in the action explains only part of it. Across states, the policy's actions span a curved region, so an affine model fits policy data only as the tangent at the policy's action, the operating point: on G1, the MLP's tangent there is as accurate as the MLP; at the mean action, 31 times worse. The offset must encode, from the state alone, where the policy acts, whereas an MLP reads the action. Supplying the policy's action removes most of the remaining gap on G1, Ant and Humanoid. The only tested online repair approaching the MLP feeds the action into the nonlinear lift, abandoning the structure. For this use, the MLP world model remains the better choice.
