Digit stabilize¶

An Agility Digit humanoid must stay upright against a worst-case torso force — a two-player avoid game (ISAACS proper): the control policy maximizes the safety value, an adversary applies the most destabilizing admissible force, and the learned value is the robust-invariance certificate.

digit stabilize rollout

Requires a custom mjlab fork to simulate

The Digit assets and cfgs are vendored and import cleanly on stock mjlab (cfg construction and list_tasks() work), but stepping the simulation needs the lab's MjlabSafety_Digit fork for its ball-joint entity patch. Until that is upstreamed, treat Digit as partially supported. See requirements.

Tasks¶

task	objective	learner (1P → 2P)
`digit_stabilize`	stand / track vs adversarial torso force (reach-avoid, real target)	`ReachAvoidPPO1P` → `ReachAvoidPPO2P`
`digit_stabilize_avoid`	stay upright forever (avoid)	`SafetyPPO1P` → `SafetyPPO2P`
`digit_stabilize_stay` / `digit_box_stabilize_*`	stance-set viability variants	`SafetyPPO1P` → `SafetyPPO2P`

(Names are the on-policy family; --family off_policy gives the *SAC* cells — note SafetyPPO2P and SafetySAC2P are structurally different games, see the MAP.)

The avoid tasks are genuine two-player avoid games — with --adversary they resolve to SafetyPPO2P (no target set). They do not emulate an avoid task with a degenerate l (the retired l_neg pattern): avoid is not a reach-avoid instance — see margins.

Margins¶

g (safety) = upright / stance integrity: torso tilt, base height, non-foot contact — negative on a fall. Digit stepping needs the MjlabSafety_Digit entity.py fork patch.
l (target, digit_stabilize only) = in the target stance set (upright, at rest). The avoid variants declare no l.

Run it¶

python examples/train.py --family on_policy --task digit_stabilize_avoid --adversary   # -> SafetyPPO2P (two-player avoid)
python examples/train.py --family on_policy --task digit_stabilize                      # reach-avoid, single-player

from robot_safety_sandbox import make_tensor, algo_name
env = make_tensor("digit_stabilize_avoid", num_envs=2048, adversary=True)
# algo_name("digit_stabilize_avoid", adversary=True) -> "SafetyPPO2P"