Bundled Contact Gradients: End-to-End Humanoid Policies in a Differentiable Simulator
Loading video
Loading videoBundled Contact Gradients (BCG) trains end-to-end humanoid policies entirely inside a differentiable simulator, bundling the gradients of contact dynamics so they backpropagate directly, removing the dependence on real hardware or a non-differentiable simulator.
Category: simulation
Author: @dyumanaditya
Date: 2026-09-28T00:00:00
Duration: 20.0s
Reference: https://arxiv.org/abs/2609.30951
