README.md

---
title: G1 Himalaya Terrain Training
emoji: 🦿
colorFrom: red
colorTo: green
sdk: gradio
sdk_version: 6.26.0
python_version: '3.11'
app_file: app.py
pinned: false
short_description: GPU training for Unitree G1 on snowy Himalayan terrain
---

# Unitree G1 — Himalayan terrain locomotion training

Trains on **real elevation data**: `dem_khumbu.npy` is SRTM-derived terrain
(AWS Terrain Tiles, 4.2 m/px, 3.2 km square) of the Khumbu glacier valley below
Everest Base Camp. `himalaya_terrain.py` cuts random 600 m crops, squeezes them
onto Playground's 20 m arena with 0.3 m of relief plus scree detail, and gives
every environment its own crop through the domain randomizer. Set `TERRAIN=playground`
to fall back to the stock 5 cm noise field; `HIMALAYA_RELIEF`, `HIMALAYA_PATCH`,
`NUM_TERRAINS` tune the terrain.

**Snow** (`SNOW=1`, default): every environment also draws its own snow —
foot–floor friction from `SNOW_FRICTION` (default `0.3,0.7`; packed snow ≈ 0.5,
rock ≈ 1.0) and a soft-contact snow depth from `SNOW_DEPTH` (default `0.0,0.08` m,
0–3 in: feet settle about half the depth at rest and up to twice it on a
landing). Randomizing both is what makes the policy robust across snow
conditions instead of tuned to one. `SNOW=0` gives bare terrain. Checkpoints go
under `himalaya-snow/` in `HF_REPO`.


Trains `G1JoystickRoughTerrain` from [MuJoCo Playground](https://playground.mujoco.org/)
with Brax PPO on the Space's GPU. Thousands of MJX environments step in parallel,
versus the 16 CPU processes a laptop manages.

**Watch a checkpoint walk.** The reward curve says whether training is moving;
the *Render rollout* section says whether the robot walks. Pick a checkpoint,
a forward speed, snow friction and depth, and it films the policy on a Himalaya
crop (`rollout.py`; MJX + `MUJOCO_GL=egl`) and reports how long it stayed up and
how far it walked. `NUM_EVALS` (default 40) sets how many evals — and therefore
log lines and checkpoint uploads — a run makes.

## Snow gait (`GAIT=snow`, the default)

Playground's joystick task is tuned for pavement: a 15 cm swing-height reference
at 1.25–1.5 Hz with the hips pinned to a narrow stance. This Space defaults to `GAIT=snow`, which retunes it
for deep snow via `snow_gait.py` — a 22 cm swing (`FOOT_HEIGHT`), a slower
1.0–1.3 Hz cadence (`GAIT_FREQ`) with more air-time reward for longer strides, a
relaxed hip-deviation cost so the stance can widen, and a new cost for letting
the feet come within `FOOT_SEPARATION` (20 cm) of each other *laterally*.

That last term is also a fix, not just styling. The joystick task ends an episode
when the two feet touch, and the 200M-step run plateaued because the gait crossed
its feet after ~45 steps and took the −100 termination every episode while still
upright. The separation cost penalises exactly that, and is measured along the
pelvis's y-axis so feet may still pass fore-aft as walking requires.

Checkpoints go under a `-snowgait` suffix.

## Fall recovery (`TASK=getup`)

Walking and getting up are different tasks, and the joystick task cannot learn
the second: it ends the episode the moment the torso tips past horizontal (with
a −100 reward), so the policy never sees a timestep while fallen. `TASK=getup`
switches the Space to `g1_getup.py`, a humanoid port of Playground's `Go1Getup`:
60% of episodes *start* fallen (dropped from 0.5 m with a random orientation and
random joint angles, then left to settle), nothing terminates on being down, and
the reward asks for an upright torso at standing height — then, only once both
hold, for the nominal pose and for the policy to stop moving.

Playground's G1 is a *feet-only* collision model, so a fallen torso would drop
through the floor. `g1_getup_model.py` rewrites the robot XML: it enables the
thigh, shin and hand collision geoms the model already carries but leaves off,
and adds torso, pelvis and head capsules sized from the visual meshes. Body geoms
get `contype=2 / conaffinity=1` so they collide with the ground but never with
each other — without that the pelvis capsule overlaps the thighs and the standing
robot is blown over by its own contacts.

Observations (93) and actions (29) keep the joystick task's layout, so a fall
detector can hand control between a walking policy and a get-up policy at run
time. Checkpoints go under a `-getup` prefix. Flat ground only: `TERRAIN` and
`SNOW` are ignored for this task.

```
TASK=getup      # then Start; ~100M steps
```

Training starts automatically on boot. Checkpoints are written to `/data` when
persistent storage is attached, and pushed to `HF_REPO` if that variable and a
write-scoped `HF_TOKEN` secret are set.

**Settings that matter**
- *Sleep time* → **never**. The Space idles on HTTP traffic, not CPU/GPU load, so
  a background training thread will not keep it awake.
- *Persistent storage* — without it, checkpoints are lost on restart. Set
  `HF_REPO` so they are pushed to the Hub instead.
- `NUM_TIMESTEPS` (default 200M), `ENV_NAME`, `SEED` are configurable variables.