Challenges the widespread assumption that auxiliary dynamics supervision creates useful latent structures for robotics.
March 24, 2026
Original Paper
Evaluating Factor-Wise Auxiliary Dynamics Supervision for Latent Structure and Robustness in Simulated Humanoid Locomotion
arXiv · 2603.21268
The Takeaway
The study shows that per-factor auxiliary losses in humanoid locomotion fail to produce decodable or disentangled latent spaces, with standard LSTMs actually outperforming complex supervised Transformers. This suggests current architectural trends in 'informed' world models for robotics may be adding complexity without functional benefit.
From the abstract
We evaluate whether factor-wise auxiliary dynamics supervision produces useful latent structure or improved robustness in simulated humanoid locomotion. DynaMITE -- a transformer encoder with a factored 24-d latent trained by per-factor auxiliary losses during proximal policy optimization (PPO) -- is compared against Long Short-Term Memory (LSTM), plain Transformer, and Multilayer Perceptron (MLP) baselines on a Unitree G1 humanoid across four Isaac Lab tasks. The supervised latent shows no evid