Porting a G1 walking recipe to the Booster K1
The TactileStep rebuild ran on the Unitree G1. The sensing-sole work runs on the smaller Booster K1. The first step toward tactile walking on the K1 was to train an ordinary walking policy there with the Hiking in the Wild recipe. Three attempts; none produced a policy that walks well in both simulators.
In short
- Two attempts from scratch learned to stand, not walk: 0.01 and 0.09 m/s at a command of 0.5 m/s after 400 iterations. Reward settings sized for the larger G1 made standing still pay.
- A warm start from my own K1 walking policy walked. At iteration 200 it passed the Isaac Lab check: no falls in 24 runs, 0.41 m/s.
- The same policy failed in MuJoCo: 10 falls in 24 runs and 0.26 m/s.
- More training made it worse. Its speed fell from 0.41 to 0.19 m/s by iteration 1,300, as the terrain curriculum got harder and the action noise grew. Training stopped at iteration 1,478.
- So the tactile stage on the K1 was not started.
What had to change from the G1 version
Hiking in the Wild is a published recipe for training a G1 walking policy on rough terrain in Isaac Lab, with a motion prior (AMP) that rewards a human-like style. Porting it to the K1 meant changing:
- The robot. The K1 simulation model, its 22 joints, new left-right mirror maps, and the manufacturer's walking gains (legs 100 / 2, ankles 50 / 1, arms 20 / 2, head 4 / 1), joint limits and default pose, with action scale 0.25.
- Sizes. Sole 0.038 m below the foot frame, feet 0.10 m apart, height scanners at the sole centre, the depth camera on the K1's head, 15° down.
- Speeds. Commands 0.3 to 0.7 m/s instead of 0.45 to 1.0. From attempt 2 on, the speed-tracking width was cut from 0.5 to 0.3 and two gait thresholds (a penalty for standing when told to walk, and a reward for time in the air) were scaled by 0.6.
- Motion data for the style prior. My own K1 clips of a catwalk-style walk, converted without retargeting (attempt 1 also used five clips that are almost in place).
- Falls. A run ends when the trunk drops below 0.30 m above its origin (attempts 1 and 2), or below 0.25 m above the lower foot (attempt 3; nominal 0.51 m).
- Terrain. Flat ground and low steps (3 to 8 cm) instead of the full parkour set.
- Attempt 3 only: the policy network starts from my own K1 walking policy (690-32-32-20), widened to 256-128 without changing its output, action noise 0.4 at the start. It needs that policy's 10 frames of history: with the recipe's 8 frames the copied policy fell in 24 of 24 runs; with 10 it reproduces the original exactly.
- Budget. At most 2,000 iterations for attempts 1 and 2, 6,000 for attempt 3; about 7.2 to 7.8 s per iteration.
Three attempts
Training-log values at iterations 50 / 200 / 400 to 500 / last. The test: command 0.5 m/s on flat ground, 3 seeds × 8 robots, 20 s.
| Attempt 1: from scratch | Attempt 2: from scratch, rescaled terms | Attempt 3: warm start | |
|---|---|---|---|
| Test speed (m/s) and falls | it200 0.00, it400 0.01; 0 falls | it200 0.00, it400 0.09; 0 falls | it0 0.43 (5/24 falls), it200 0.41 (0/24, pass), it500 0.28, it1300 0.19 |
| Episode length (steps) | 34 / 41 / 33 / 112 | 37 / 67 / 70 / 63 | 140 / 619 / 580 / 769 (it1478) |
| Speed-tracking reward per episode | 0.78 / 1.50 / 1.61 / 1.47 | 0.47 / 1.06 / 1.18 / 1.19 | 1.36 / 13.5 / 6.5 / 13.3 |
| Curriculum tracking score | 0.02 to 0.04 | 0.01 to 0.03 | 0.03 / 0.34 / 0.16 / 0.33 |
| Air-time reward / standing penalty | 0.003 / −0.19 | 0.003 / −0.13 to −0.21 | 0.01 to 0.05 / −0.7 / −1.1 / −2.2 / −1.4 |
| Style (AMP) reward | 0.32 → 0.50 | 0.30 → 0.45 | 0.30 / 0.34 / 0.37 / 0.42 |
| Action noise (std) | 0.92 → 0.87 | 0.91 → 0.85 | 0.40 / 0.43 / 0.49 / 0.65 |
| Terrain level | 0 → 3.8 | 0 → 3.0 | 0.08 / 0.82 / 4.17 / 3.10 |
| Ends: trunk too low / time-out | 0.89 → 0.32 / 0 → 0.63 | 0.86 → 0.35 / 0 → 0.62 | 0.58 / 0.26 / 0.26 / 0.20; time-out 0.17 → 0.78 |
Reading the table. Attempts 1 and 2 never learned to move: short episodes, a flat tracking reward and a test speed near zero. Attempt 3 started from a policy that already walks and kept walking, with long episodes and a much higher tracking reward. Its test speed, though, peaked at iteration 200 and then fell.
Why the first two stood still
Attempt 1 was stopped at iteration 558. Its checkpoint at 500 moved 0.01 m/s at a command of 0.3, 0.00 at 0.5 and 0.22 at 0.8; with action noise on, 0.19 at 0.5. It did not fall. Since it did move at 0.8, the test was working; the cause lay in the port:
- The commands were scaled down to the K1, but the tracking width of 0.5 stayed at the G1 value. A robot standing still at a command of 0.5 m/s therefore kept 37 % of the tracking reward (the G1 about 12 % at its mean command).
- The penalty for standing when told to walk was still sized for the G1.
- 38 % of the style-prior samples came from the clips that are almost in place.
Attempt 2 fixed these (tracking width 0.3, thresholds × 0.6, only the catwalk clips) and still reached only 0.09 m/s at iteration 400, with the tracking reward flat at 1.06 to 1.19 per episode. After two failures from scratch, the third attempt used a warm start.
Why the third got worse
Read from the training numbers only, without a separate experiment: between iterations 200 and 500 the terrain level jumped from 0.8 to 4.2 (onto the low steps), the tracking reward halved (13.5 → 6.5) and the standing penalty doubled (−1.1 → −2.2). The action noise kept rising (0.40 → 0.65, driven by the entropy bonus) and the style reward kept rising (0.30 → 0.42), while the speed of the noise-free policy on flat ground fell from 0.41 to 0.19 m/s. By iteration 1,478 the tracking reward in training had recovered to 13.3: the noisy policy still tracked during training, but its mean action walked slower.
| Attempt 3, iteration | 200 | 500 | 700 | 1,000 | 1,300 |
|---|---|---|---|---|---|
| Test speed at command 0.5 (m/s) | 0.41 | 0.28 | 0.22 | 0.21 | 0.19 |
Reading the table. The best policy came early and training only took it further away. The decay coincides with both the harder terrain and the growing action noise; which of the two causes it was not separated.
The best checkpoint in Isaac Lab and MuJoCo
Attempt 3 at iteration 200, tested in the training simulator (Isaac Lab) and in MuJoCo with the same observations and depth pipeline (training friction 1.0, the same PD gains applied as torque control, the robot's effort limits, no armature; 3 seeds × 8 perturbed starts, 20 s). Command 0.5 m/s.
| Simulator | Falls | Speed (m/s) | Heel / flat / toe | Peak landing force | Check |
|---|---|---|---|---|---|
| Isaac Lab, iteration 0 (the starting policy) | 5 / 24 | 0.43 | 0.45 / 0.54 / 0.00 | 255 N | – |
| Isaac Lab, iteration 200 | 0 / 24 | 0.41 ± 0.00 | 0.23 / 0.76 / 0.01 | 297 ± 6 N (1.54 BW) | pass |
| MuJoCo, iteration 200 | 10 / 24 | 0.26 ± 0.01 | 0.29 / 0.60 / 0.11 | 262 ± 4 N (1.36 BW) | fail |
Reading the table. In Isaac Lab, 200 iterations of the recipe removed the starting policy's falls at about the same speed and shifted its landings from heel toward flat. In MuJoCo the same policy fell in 10 of 24 runs (5, 3 and 2 per seed) and walked at 0.26 m/s, so it does not transfer between simulators. BW is body weight. The MuJoCo replica was validated beforehand: at the first step its joint readings match Isaac Lab up to noise and its depth image is identical.
Clips
Training recipe: Hiking in the Wild (CC BY-NC 4.0), used for research. The policies shown were trained on the K1 here; the pretrained G1 policy is not shown.
What to try next, and limits
- Keep the warm start, hold the curriculum. Stay on flat ground until the test speed holds, and fix the action noise at 0.4 with no entropy growth: the decay coincides with both.
- Commands. The test gives a constant 0.5 m/s, while training uses commands toward targets that slow down near them. Add episodes with constant speed.
- Simulator transfer. Test the starting K1 policy itself in MuJoCo first.
- Limits. Three attempts, one recipe, short budgets, simulation only. No sensing sole was involved yet, and no course was run.
Scores here are not F1: test speed at a command of 0.5 m/s, falls out of 24 runs (3 seeds × 8 robots, 20 s), share of heel, flat and toe landings, and peak landing force. Training recipe: Hiking in the Wild (CC BY-NC 4.0), used for research; its code, data and weights are not included. Related: TactileStep on the Unitree G1, the simulated sole up close, step 1, first simulation run.