Verified, then logged.
Sun 22 Nov 2026 · 00:46 KST / 15:46 UTC한국어 | ENSearch /

01 / AI · Analysis

E-prop posts 100% survival in a humanoid robot walking trial

A learning rule modelled on biological nervous systems matched the survival rate of a non-biological baseline.

Monitors at an unoccupied computational physics workstation show heat diffusion and a steepening nonlinear wave alongside subtly differing numerical simulations. AI-generated image

Researchers reported that the e-prop learning method achieved a 100% survival rate in a humanoid walking experiment.

The team compared learning rules to see whether rules modelled on biological nervous systems could sustain walking performance.

The paper’s abstract compared survival rates across the methods and described a short stepping demonstration on a physical robot.

Survival rates in the humanoid walking test

e-prop100
Backpropagation through time100
Source arxiv.org · as of 2026-11-20 · unit: %Chart: made by LOGIT

Survival rates across training methods

The researchers used backpropagation through time as the non-biological baseline. They reported that e-prop matched its survival rate.

Across all methods, the team trained the neural network to imitate a control policy learned through reinforcement learning.

According to the abstract, one method kept the network’s recurrent connections fixed and trained only the linear output layer. The method that adjusted connections based on spike timing and rewards had the lowest survival rate.

Two-legged stepping on a full-size robot

The researchers reported that they demonstrated two-legged stepping on a full-size Unitree H1-2 humanoid robot.

The abstract described the stepping as brief and gave neither the duration of the demonstration nor the number of steps.

Jihwan Nam (AI-written)