Go2 crawl

A Unitree Go2 must duck under a low bar or stop — the campaign's second benchmark for temporal commitment: a closing gate creates a region where standing still eventually becomes unsafe, so the robot must decide to crawl through in time. Unlike the gap, there is no gait-phase launch conversion, which sidesteps the co-adaptation wall.

Experimental — a folded thread

The crawl / low-bar / tunnel line is shelved: a model-free reach-avoid executor hit a ~0.3 success ceiling on sustained crawling (a sustained crouch-crawl is a harder problem than the one-decision gap jump). The environments below remain registered for reference but are not recommended entry points and may not reproduce a clean result. The negative finding — and the filter-as-gate pivot for resuming it — is preserved in the project record.

Tasks

task	objective	learner
go2_crawl	duck under the bar → safe rest past it	ReachAvoidPPO1P
go2_crawl_duck	momentum approach at a low bar + forward-velocity reach	ReachAvoidPPO1P
go2_crawl_gate_ra / _gate_avoid	closing-gate twins: descending virtual ceiling (RA vs avoid)	ReachAvoidPPO1P / SafetyPPO1P
go2_crawl_twin_ra / _twin_avoid	static-bar twins (negative control: no committed region ⟹ avoid == RA)	ReachAvoidPPO1P / SafetyPPO1P
go2_crawl_isaacs	crawl + worst-case base-force adversary	ReachAvoidPPO2P

The *_twin_* and *_gate_* pairs are the claim's controlled experiment: the avoid and reach-avoid twins share one g, and differ only in whether an l target is present — the contrast should appear only where a committed region exists (the closing gate), not on the static bar.

Margins

• g (safety) = crawl integrity: terrain-relative height / tilt / non-foot contact, plus (gate variant) a virtual descending-ceiling term and a crushed_by_gate termination.
• l (target) = forward-velocity liveness / rest past the bar (≥ 0 once through). The avoid twin declares no l (compose(g_fn)); see margins.

Run it

python examples/train.py --family on_policy --task go2_crawl_gate_ra        # reach-avoid, closing gate
python examples/train.py --family on_policy --task go2_crawl_gate_avoid     # avoid control

from robot_safety_sandbox import make_tensor
env = make_tensor("go2_crawl_gate_ra", num_envs=2048)