$\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation
$\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation
This paper introduces $\nabla$Mimic and NablaEngine, a differentiable GPU simulation engine for the Unitree G1. The authors address the limitations of zeroth-order methods like PPO and unstable first-order approaches by using exact simulator gradients bounded by gradient clipping and multistep regression. NablaEngine computes MuJoCo's full contact dynamics and adjoints as fused megakernels, achieving analytic gradients that match float64 finite differences and are 11.5 times faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, $\nabla$Mimic reaches a score of 0.82 in 20 minutes (matching the state-of-the-art world-model method) and 0.874 in 1.8 hours, whereas the standard first-order baseline SHAC stops due to exploding gradients within minutes. The code is released as open-source infrastructure.
The authors present a new differentiable simulation engine and policy learning method that achieves state-of-the-art humanoid motion tracking scores in significantly reduced time compared to zeroth-order baselines, while explicitly addressing gradient explosion issues inherent to contact-rich dynamics.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S2
Reported numbers
Reported duration
20 minutes
View original evidence
ll contact dynamics with fused per-world kernels and the adjoint of a whole control step as a single megakernel, with analytic gradients that match float64 finite differences, $11.5\times$ faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, $\nabla$Mimic reaches the score reported by the state-of-the-art world-model method, 0.82, in 20 minutes inste
Open source S5speedup_factor
11.5
View original evidence
ll contact dynamics with fused per-world kernels and the adjoint of a whole control step as a single megakernel, with analytic gradients that match float64 finite differences, $11.5\times$ faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, $\nabla$Mimic reaches the score reported by the state-of-the-art world-model method, 0.82, in 20 minutes inste
Open source S5
- code: not_reported
- paper: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2610.41654v1 Announce Type: new Abstract: Robot policy learning is dominated by zeroth-order methods such as PPO, which treat the simulator as a black box and pay for it with enormous sample counts. First-order policy learning differentiates through the dynamics, but on contact-rich whole-body control it has rarely worked: its gradients explode. A recent world-model method showed that derivative information can
Open source S2
ll contact dynamics with fused per-world kernels and the adjoint of a whole control step as a single megakernel, with analytic gradients that match float64 finite differences, $11.5\times$ faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, $\nabla$Mimic reaches the score reported by the state-of-the-art world-model method, 0.82, in 20 minutes inste
Open source S5
ontact solve. $\nabla$Mimic bounds these 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, which bounds every action target independently of the gradient scale. NablaEngine, a differentiable GPU simulation engine for the Unitree G1, removes the cost: it computes MuJoCo's fu
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org