Exact simulator gradients made first-order humanoid policy learning practical, and the differentiable engine behind it was 11.5 times faster than the available MuJoCo Warp adjoint
AspireAI community · arXiv
Robot policy learning is dominated by zeroth-order methods such as PPO that treat the simulator as a black box and pay in sample count. First-order policy learning differentiates through the dynamics, but on contact-rich whole-body control its gradients explode, and a recent world-model method got around this using surrogate gradients of a learned model. The authors show exact gradients of the simulator itself do better once two obstacles are removed: exact gradients contain rare unbounded samples because contact dynamics is not differentiable everywhere and its derivative diverges near some events, and they are expensive because the simulator's adjoint must replay the contact solve. NablaMimic bounds the samples with median-relative per-sample gradient clipping at every step of the backward pass and reuses each costly simulated window by multistep regression onto gradient-derived action targets, bounding every action target independently of gradient scale. NablaEngine, a differentiable GPU simulation engine for the Unitree G1, computes MuJoCo's full contact dynamics with fused per-world kernels and the adjoint of a whole control step as a single megakernel, with analytic gradients matching float64 finite differences, 11.5x faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU it reaches the score reported by the state-of-the-art world-model method, 0.82, in 20 minutes instead of 1.2 hours, and 0.874 in 1.8 hours, while the standard first-order baseline SHAC stops on exploding gradients within minutes. NablaEngine and NablaMimic are released as open-source infrastructure. The limitation is that the results are on the Unitree G1 with MuJoCo-family contact dynamics, so the gradient-stability argument has not been shown for other simulators or for non-humanoid embodiments.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.