Walking with a sensing sole: TactileStep on the Unitree G1
The other pages use the sole to read the floor. This one asks whether it can change the walk. TactileStep trains a humanoid walking policy that feels its soles and is rewarded for landing softly and using more of the sole. I rebuilt it on the Unitree G1, starting from the pretrained Hiking in the Wild policy, and compared the two at the same walking speed. The new policy did not land softer.
In short
- Built: TactileStep's 60-point sole per foot, its training phases and its tactile rewards, on top of the Hiking in the Wild G1 task, trained from that policy for 2,000 iterations.
- An early pass was an artefact of speed. At iteration 400 the new policy landed with 600 N against 680 N, but it walked at 0.61 m/s instead of 0.80. Against the pretrained policy at the same speed it lost: 600 against 534 N.
- Speed-matched, every checkpoint fails. The final policy lands at 782 ± 10 N against 680 ± 20 N at 0.80 m/s, and uses no more of the sole (0.944 against 0.946).
- The gait changed. Early checkpoints land toe-first on every step; the final one lands 23 % heel, 27 % flat, 50 % toe, against 0 / 58 / 42 for the pretrained policy.
- Little room to gain on flat ground. The pretrained policy already uses about 95 % of the sole.
What was built
TactileStep is a published method for humanoid walking with tactile soles. The policy sees a pressure picture of each sole; extra rewards ask for a low impact at landing, a wide contact area and a centre of pressure well inside the foot. Hiking in the Wild is a published G1 walking policy trained on a rough-terrain (parkour) task; its released policy was the starting point.
As in the paper
- 60 sensing points per foot (a 12 × 5 grid on the shoe sole), read by rays along the sole normal, with the paper's weighting, force projection, smoothing between neighbours, and force, area and centre-of-pressure features.
- Two frames of sole history for the policy (16 numbers); the critic also gets the foot's vertical speed and the gait phase.
- The paper's training phases, all its tactile reward weights and thresholds, and its sensor noise (±10 % per point, a small offset per episode, 8 % one-frame delay).
- Two critics with advantage mixing [1.1, 0.9], motion-prior (AMP) weight 0.25, foot-slide penalty −0.4. The Hiking rewards, terrain and motion prior stay unchanged.
Different from the paper
- Warm start from the released Hiking policy (a plain 512-256-128 network) instead of training from scratch with a mixture-of-experts network. The new sole inputs start at zero weight; the critic warms up first while the policy is frozen.
- Far shorter training: 2,000 iterations in total against the paper's 50,000. Learning rate 3e-4, initial action spread 0.3, the tactile ramp-in shortened to 200 iterations (paper: 5,000) and switched off on resume.
- Terrain: flat ground and low stairs (5 to 12 cm) for the main run, with no entropy bonus. 4,096 parallel robots.
- Reward grouping: only the new landing terms go to the second critic; the paper also moves several other terms there.
- Constants the paper does not give had to be chosen, among them the contact thresholds (force above 0.05 body weight or area above 0.15), a pre-landing height of 0.08 m and a landing window of 0.1 s. The terrain normal per sensing point is computed from the hit points, because the simulator's ray caster gives no normals.
How it was tested
- Flat ground, 3 seeds × 8 robots × 21 s, the same script and the same simulated sole for both policies.
- Three measures: peak force at each landing, the share of the sole in contact during stance, and whether the foot lands heel-first, flat or toe-first. Falls are counted.
- Pass rule: lower landing force and larger contact area than the pretrained policy, each by more than twice the spread over seeds, with no falls.
- Same speed. The new policy is commanded 0.7 m/s. The pretrained policy is run at whichever command gives the same actual speed, within 0.02 m/s. Its commands 0.5, 0.631, 0.665 and 0.7 give 0.61, 0.74, 0.77 and 0.80 m/s.
The early pass and why it was wrong
A first run on the full Hiking terrain curriculum walked, but worse than the pretrained policy (iteration 400: 834 ± 41 against 680 ± 20 N, contact area 0.820 against 0.946, 0.67 m/s). The main run then trained on flat ground and low stairs, with a check every 30 minutes that stopped training on a pass. It passed at iteration 400.
| Iteration 400 against the pretrained policy | Actual speed (m/s) | Peak landing force (N) | Stance contact area | Heel / flat / toe | Check |
|---|---|---|---|---|---|
| New policy, command 0.7 | 0.611 ± 0.003 | 600 ± 27 | 0.951 ± 0.001 | 0.00 / 0.00 / 1.00 | – |
| Pretrained, same command 0.7 | 0.803 ± 0.002 | 680 ± 20 | 0.946 ± 0.002 | 0.00 / 0.58 / 0.42 | pass |
| Pretrained, command 0.5: same speed | 0.610 ± 0.002 | 534 ± 15 | 0.960 ± 0.001 | 0.01 / 0.69 / 0.30 | fail |
Reading the table. Against the pretrained policy at the same command, the new one looked better: 80 N lower force and slightly more contact. But it walked at 0.61 m/s instead of 0.80, and slower walking lands softer by itself. Run at 0.61 m/s, the pretrained policy lands with 534 N and 0.960 contact area, so the new policy is worse on both (+66 N, −0.009). The pass was the speed, not the sole. Every later check matched speeds.
Speed-matched, up to 2,000 iterations
Training resumed from iteration 400 to 2,000. Each checkpoint was compared with the pretrained policy at the same actual speed. Neither policy fell in any test (0 of 24 robots).
| Checkpoint (actual speed) | Pretrained command → speed | Peak force, new vs pretrained (N) | Contact area, new vs pretrained | Heel / flat / toe, new | Check |
|---|---|---|---|---|---|
| 400 (0.61 m/s) | 0.5 → 0.61 | 600 ± 27 vs 534 ± 15 | 0.951 ± 0.001 vs 0.960 ± 0.001 | 0.00 / 0.00 / 1.00 | fail |
| 800 (0.74), closest | 0.631 → 0.74 | 707 ± 11 vs 647 ± 6 | 0.949 ± 0.001 vs 0.953 ± 0.001 | 0.00 / 0.00 / 1.00 | fail |
| 1200 (0.77) | 0.665 → 0.77 | 826 ± 34 vs 673 ± 3 | 0.947 ± 0.001 vs 0.949 ± 0.001 | 0.00 / 0.14 / 0.86 | fail |
| 1700 (0.79) | 0.7 → 0.80 | 835 ± 24 vs 680 ± 20 | 0.944 ± 0.001 vs 0.946 ± 0.002 | 0.02 / 0.48 / 0.50 | fail |
| 2000, final (0.80) | 0.7 → 0.80 | 782 ± 10 vs 680 ± 20 | 0.944 ± 0.000 vs 0.946 ± 0.002 | 0.23 / 0.27 / 0.50 | fail |
Reading the table. At every checkpoint the new policy lands harder than the pretrained one at the same speed, by 60 to 155 N, more than twice the spread over seeds. Its contact area is equal or slightly lower (differences of 0.002 to 0.009). The closest checkpoint is 800, with the smallest force excess (+60 N). In body weights, the final policy lands at 2.391 against 2.080. With more training the policy regained the speed (0.80 m/s at iteration 2,000) and its landings moved from all toe-first to a mix with some heel strikes; the foot's pitch at touchdown went from −9.3° at iteration 400 back to −2.3°, close to the pretrained −2.4°.
The feet, side by side
Left: the pretrained Hiking in the Wild policy. Right: TactileStep. Both commanded 0.7 m/s on flat ground, so in these clips the two do not walk at the same actual speed.
Pretrained policy: Hiking in the Wild (CC BY-NC 4.0), shown for research.
What it means and its limits
- The rebuild did not reproduce the benefit. Within 2,000 iterations the soft-landing terms did not lower the peak force, and the area term did not widen the contact.
- Little room on flat ground. The pretrained policy already uses about 95 % of the sole on flat ground. The paper reports a gain from 0.912 to 0.929; here the starting point is higher.
- Not the paper's training. A warm start, a smaller plain network, 2,000 instead of 50,000 iterations, a shortened tactile ramp-in, several chosen constants, and flat ground with low stairs only. A longer run or a run from scratch could behave differently; this one did not test that.
- Speed matters for every comparison of landing force. The early pass shows how easily a slower gait looks like a softer one.
- Simulation only, with a simulated sole and no real robot.
Scores here are not F1: peak landing force in newtons, stance contact-area ratio, and the share of heel, flat and toe landings, each mean ± sd over 3 seeds of 8 robots on flat ground. Pretrained policy: Hiking in the Wild (CC BY-NC 4.0), shown for research; its code, data and weights are not included. Related: porting the Hiking recipe to the Booster K1, the simulated sole up close, step 1, second simulation run.