Add jump task: two-foot hop, land standing

[SAMBAS123]

Why

Roulade is a roll. Kick is a swing. Sitstand is a sit. Nothing in the pad means leave the ground and land — swap the ONNX in, the duck hops, then auto-hands back to stand/walk after ~2 s.

A two-foot hop is at the XL330 margin (open-loop BAM crouch→extend never left the floor). The env therefore pays Δpeak height and Δbilateral flight only while both feet are off, upright, and at/above stand — a sit/fall with feet off does not count. Reverse-curriculum airborne / crouch-upward spawns put landing on-policy even if takeoff is rare.

What

• New env Mjlab-Jump-Flat-MicroDuck (+ backlash twin on the allcollisions model)
• 61D actor obs: 48 proprio + twist(3) + head(4) pad + body(6) pad — hot-swappable with walk/stand/roulade
• Reset mix: 60% standing (learn takeoff) / 25% crouched with upward vz / 15% already airborne (learn the landing); curriculum shifts toward standing starts
• scripts/infer_policy.py --jump (key J). Auto-return after 2 s — jump ends standing
• CPU cfg tests lock the contract, signs, sit-vs-hop gates, and registration

AGENTS.md bar

uv run --with pytest pytest tests/test_jump_cfg.py
uv run train Mjlab-Jump-Flat-MicroDuck --env.scene.num-envs 64 --agent.max_iterations 5

Ran on a Spark (GB10): 16 jump tests green. Smoke 64 envs × 5 iters: actor obs (61,), every reward term computed, Episode_Termination/nan_state = 0, self-negating penalties logged ≤ 0.

This is the env. A real run is ~2000–3000 iters at 4096 envs (experiment_name=microduck_jump), then scripts/export.py (bakes the normalizer — mandatory). A local 4096-env Spark run produced a visible ~3 cm hop from a standing start; weights stay out of git (same as the other tricks).

Deploy rehearsal

uv run scripts/infer_policy.py --walking walk.onnx --standing stand.onnx \
    --jump jump.onnx --new-cmd-obs

J hops; after ~2 s control hands back to standing/walking.

[abhinav3]

Hi, have you tested that it is able to jump with its tiny motors? Could you pls share the simulation video?