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.
TeamZ trained a G1 policy in Isaac Sim and tested it on uneven ground. Three terrains produced 20-second runs covering about 39–40 m, with confirmed intervals when both feet were off the ground followed by alternating landings. The same model drifted out of the test area at a lower speed command. These are simulation results, not hardware tests.
Video: commanded speed 2.2 m/s, measured mean about 2.0 m/s, forward travel 39.45 m in 20 seconds. Normal simulation speed, 50 fps. Warehouse shelves are background scenery; the ground under the feet has physical height variation.
What we trained
We used Isaac Sim 5.1 and an Isaac Lab G1 environment with a 37-joint model. TeamZ trained a control policy with PPO on the existing simulation and learning framework, rather than simply running a publicly pretrained policy. Training mixed flat ground with rough terrain having height variation from −2 to +2 cm. We progressively adjusted velocity tracking, posture and leg-lift timing, completing 3,200 training updates.
Inputs include terrain heights and body velocity obtained from the simulator. Evaluation also uses simulator orientation for heading correction. This is a locomotion-control test, not navigation based on cameras or Visual SLAM.
Staying upright and moving fast were not enough
The initial model moved for 20 seconds but advanced only about 30 cm: it circled in place. Body-frame forward velocity could miss that failure, so we added world-frame displacement and lateral drift to evaluation.
Increasing a turning penalty produced early falls. Several rewards changed together, so we cannot attribute the outcome to one cause. We revised settings, including capping penalties, and trained again. The next model advanced about 30 m, but its rapid foot switching did not match the intended gait.
Intermediate model, 1.6 m/s command: it remained active for 20 seconds but left the rough-terrain test area partway through. We do not count this as a completed rough-terrain run.
We then added a 0.6-second leg-switching target, swing-foot lift targets and timing for unloading both feet. Gait changes also affected heading tracking, so we checked displacement and drift after each training iteration.
The final model drifted at the lower command
At 1.6 m/s, the final model was tested in three flat-ground and three rough-terrain trials. All six drifted sideways and left the test area after about 12–15 seconds.
Final model, terrain seed 102: the video ends when it leaves the area at 15.22 seconds. Reset sequences have not been joined together.
We then changed the command to 2.2 m/s, within the training speed range. All three rough-terrain trials continued for 20 seconds and remained in the test area.
| Terrain seed | Duration | Forward travel | Final lateral drift | Mean speed | Confirmed flight intervals |
|---|---|---|---|---|---|
| 100 | 20 s | 40.00 m | 0.37 m | 2.044 m/s | 31 |
| 101 | 20 s | 38.81 m | 2.35 m | 1.993 m/s | 37 |
| 102 | 20 s | 39.45 m | 0.24 m | 2.005 m/s | 34 |
Mean speed is world-frame X velocity after the first three seconds. Flight counts also exclude those first three seconds. Seeds specify random generation, including terrain. A lower speed command was not necessarily easier. Additional comparisons are needed before attributing success solely to the higher speed. This exploratory experiment added conditions in response to intermediate results, using one trained policy and three terrains; it does not establish broad reliability.
How we checked that both feet left the ground
A momentary reduction in contact force was not sufficient evidence. We recorded foot forces, positions and orientations at 200 Hz and separately calculated foot collision geometry and local terrain height.
We counted intervals lasting at least 20 ms when both foot contact forces were below threshold and both collision shapes were conservatively estimated to be at least 1 cm above the local ground. Thresholds of 0.5 N, 1 N and 5 N produced the same counts. Landing order alternated left and right.
This conservative rule does not detect every flight interval. It confirms separation from the ground for qualifying intervals and provides evidence of repeated flight phases in addition to the video.
What remains before practical use
The test strip was 8 m wide, and one trial ended with 2.35 m of lateral drift. This does not show straight running through a narrow factory aisle. Low-speed transitions, stopping, turning, larger steps and different surfaces remain to be tested.
Results use a 37-joint model with simplified collision geometry. They have not been reproduced on a physical G1 and do not include terrain perception for autonomous navigation. Next we plan to evaluate heading control across speed changes and connections to perception and localization. Recording failure conditions as well as distance helps identify the next implementation work.
Primary sources and execution conditions
Isaac Sim 5.1; Isaac Lab at commit 3c6e67bb5c7ada942a6d1884ab69338f57596f77; RSL-RL 3.0.1. TeamZ used the existing G1 environment and PPO framework and implemented training settings, evaluation and recording.
This article publishes our evaluation results and simulation footage. It does not redistribute third-party code, weights, robot USD files or warehouse USD assets.