Samonte/g1-run-jump-policies

Unitree G1 run–jump policy stack

Frozen policies for task-humanoid-run-jump: BeyondMimic tracker → AMP run → AMP jump → hierarchical PPO switcher.

Use these files, not best_agent.pt from any training log.

Layout

Path	What	I/O
tracker/tracker.pt	TorchScript BeyondMimic 1-frame tracker (epoch 11200)	157-D → 29-D joint PD
run/policy.pt	TorchScript AMP run actor	134-D → 64-D target frame
jump/policy.pt	TorchScript AMP jump actor (stay-up)	156-D → 64-D target frame
hl/best.pt	skrl PPO HL switcher (agent_345000.pt)	116-D → 6-D [gate, vx, vy, ωz, a_h, a_flight]

skrl checkpoints (run/agent_300000.pt, jump/agent_400000.pt, hl/agent_*.pt) are for resume / play.py, not TorchScript. tracker/epoch_11200_score.ckpt resumes training in humanoid-g1-tracking.

Recommended deploy set

frozen_policies/tracker.pt     ← tracker/tracker.pt
best_policy/run/policy.pt      ← run/policy.pt
best_policy/jump/policy.pt     ← jump/policy.pt

HL (play / eval):

cd /workspace/humanoid/task-humanoid-run-jump
/venv/env_isaaclab/bin/python scripts/skrl/play.py \
  --task Nepher-G1-RunJumpHL-Play-v0 --algorithm PPO --headless \
  --checkpoint hf_policies/hl/best.pt

Why these checkpoints

• Tracker: g1_bm_l2c2_1frame_hurdle epoch 11200, future_steps=[1]. Same bytes as humanoid-g1-tracking/.../exported/policy.pt.
• Run AMP: g1_run_amp/2026-09-09_04-58-38 agent_300000. JIT already installed as best_policy/run/policy.pt.
• Jump AMP: g1_jump_amp/2026-09-09_18-11-29 agent_400000. Stay-up hop-and-glide (~0.5–0.9 m). Not a physical-bar hurdle clearer.
• HL PPO: stay-up run g1_runjump_hl_ppo/2026-09-11_05-33-19 agent_345000 — peak ~47% finish on easy 1-hurdle stage 0 (0.35–0.55 m). Alternate last keeper: hl/agent_410000.pt (~45–46%). Later resumes (tight vy/wz, stage-1 relock) did not beat this.

Honest limits

• Jump actor does not clear 0.9–1.2 m cuboids. EnvHub miner scores (~0.77) need a real hurdle-clearing jump policy.
• HL was not EnvHub-eval’d. Training finish ~45–47% on a 1-hurdle procedural course is not a 0.77 navigation score.
• best_agent.pt in skrl logs is not a keeper (wrong metric / wrong step).

License

BSD 3-Clause (same as the training repos).