Original experiment footage is shared with the Japanese edition. Some on-screen labels are in Japanese; methods, results and video descriptions are provided in English below.
In-house simulation research, September 7, 2026. We connected a public G1 running policy to Isaac Sim. The first test fell in less than a second. After checking how the initial pose related to the running phase, we verified 30 seconds on flat ground and then 60 seconds in a warehouse with shelves and cargo.
Who developed the controller?
The pretrained model is from Chasing Autonomy, published March 26, 2026 by Caltech’s Zachary Olkin, William D. Compton, Ryan M. Bena and Aaron D. Ames. It is available through Zolkin1/robot_rl on GitHub and zolkin/robot_rl on Hugging Face. TeamZ performed no additional reinforcement learning.
Paper · Authors’ code · Public models used
The model updates joint-angle targets at 50 Hz from 74 observations, including torso angular velocity and tilt, 21 joint positions and velocities, velocity commands, previous output and running phase. PD control drives the joints while Isaac Sim computes gravity, contact and friction at 200 Hz. We do not advance the torso by overwriting its position after the start.
The initial pose and phase did not match
The distributed files contain a moving pose and velocity but no corresponding initial phase. Starting at phase 0 caused a fall at 0.68 seconds. We read the authors’ validation settings and selected joint friction within the training range. Matching friction alone did not solve it.
We tried eight starting phases. Phase 0.5 sustained 12 seconds and then a separately started 30-second test. This is a reproduction condition for this initial state, not evidence that phase 0.5 works for every starting pose.
Connect long-run footage to measurements
We combined NVIDIA Simple Warehouse assets into a long aisle. To address a washed-out white robot, we transferred per-part URDF colors to USD materials and adjusted reflectance and roughness under warehouse lighting. We did not measure or reproduce the physical robot’s colors. Steering corrections use simulator ground-truth position and orientation to limit lateral drift; the video does not demonstrate Visual SLAM or Nav2 autonomy.
At 200 Hz, we record torso position and velocity, both foot contact forces and joint angles. The evaluation interval starts after three seconds. We count both feet as airborne when each resultant contact force is below 5 N, checking repeated loss of ground contact rather than labeling any fast movement as successful running.
Numerical results and implementation scope (Japanese).
Limits and what comes next
The initial condition provides a running pose and about 3.66 m/s initial velocity, then decelerates toward 2 m/s. Starting from rest and stopping are not implemented. The result applies to the specified static environment; physical robots, moving obstacles and success rates across conditions remain untested.
Next steps are start/stop behavior and integration with visual localization and route following. Fixing model interfaces, starting conditions and measurement methods provides a baseline for identifying regressions after integration.
Environment: Isaac Sim 5.1, Isaac Lab 2.3.0, NVIDIA L4. Authors’ code revision: 3c280e1. Model: running_clf_sym, public checkpoint 2026-02-27_11-09-13.
LET’S BUILD TOGETHER
What are you working on?
Tell us about the software you need to implement or test. We can discuss the work, the development setup and how we could contribute.
Discuss your development needs