Go2 stabilize and locomotion¶

A Unitree Go2 quadruped must stay upright — and, in the locomotion variant, track a velocity command — against a worst-case base force. It is the simplest and most-reproduced safety benchmark in the repo and the recommended entry point after the car-goal tutorial. The go2_stabilize reach-avoid task is the one that produced the best-ever result on this stack (a ReachAvoidSAC2P policy).

Status and requirements¶


Maturity	Stable (`go2_stabilize`, `go2_walker_flat`) · Research (`go2_locomote`)
Requires	stock mjlab sim stack (no fork); NVIDIA GPU
Adversary	supported (`go2_stabilize`, `go2_locomote`)
Recommended config	`configs/go2_stabilize_reachavoidsac2p.yaml` (SAC2P) · `configs/go2_stabilize_reachavoidppo1p.yaml` (PPO1P) · `configs/go2_walker_flat.yaml` (task policy)

Task IDs¶

Task ID	Mode	Objective	Learner (1P)	Adversary
`go2_stabilize`	reach-avoid	stand / stay upright at rest vs adversarial base force	`ReachAvoidPPO1P` / `ReachAvoidSAC1P`	✔ → `*2P`
`go2_locomote`	reach-avoid	track a forward velocity command vs adversarial force	`ReachAvoidPPO1P`	✔ → `*2P`
`go2_walker_flat`	cumulative	dense-reward flat walker — the task policy filters wrap	stock `PPO`	—

What success and failure mean¶

Failure (g < 0): the robot falls — base height below the terrain-relative floor, excessive tilt, or a non-foot body contacts the ground. The episode terminates.
Success (l ≥ 0, reach-avoid): the robot is in the target stance set (upright and, for go2_locomote, tracking the commanded velocity).

Observation and action spaces¶

Observations are proprioceptive (base orientation / angular velocity, joint positions and velocities, previous action, and the command for the locomotion variant). The two-player critic uses an asymmetric (larger) observation group.
Action: 12 joint targets (ctrl_dim=12).
Adversary action (--adversary): a 3-dim base wrench (force), scaled by the training-time magnitude; --dstb-scale sweeps it at eval.

Margins¶

g (safety) = upright / stance integrity: terrain-relative base height, tilt, and non-foot contact — negative on a fall. Rides on the reward channel; the env terminates on g < 0.
l (target) = in the target stance set (upright, at rest for go2_stabilize; tracking the command for go2_locomote).

See margins and robot_safety_sandbox/margins.py.

Training budget and hardware¶

Modest — go2_stabilize needs no curriculum or staging. The SAC2P recipe fits in ~2 GB of GPU memory (the 1M-transition replay is CPU-side) and trains in roughly an hour or two on a single modern NVIDIA GPU. go2_walker_flat is a larger dense-reward run (4096 envs, ~150M env-steps).

Train¶

# reach-avoid, single-player (PPO)
python examples/train.py --config configs/go2_stabilize_reachavoidppo1p.yaml

# reach-avoid, two-player (SAC) — the reference recipe
python examples/train.py --config configs/go2_stabilize_reachavoidsac2p.yaml

# the dense-reward task policy the filters wrap
python examples/train.py --config configs/go2_walker_flat.yaml

Evaluate¶

# the safety policy's own reach / safe rates, unfiltered
python examples/eval.py --task go2_stabilize \
    --safety-policy runs/go2_stabilize_reachavoidsac2p/final_model.zip --safety-only \
    --no-filter --num-envs 256 --steps 1000

# filter gauntlet: the walker under a swept adversarial attack
python examples/eval.py --task go2_locomote --adversary \
    --task-policy   runs/go2_walker_flat/final_model.zip \
    --safety-policy runs/go2_stabilize_reachavoidsac2p/final_model.zip \
    --filter value --dstb policy --dstb-scale 0.5 --num-envs 256

Visualize¶

python examples/play.py --task go2_stabilize --algo ReachAvoidSAC2P \
    --run runs/go2_stabilize_reachavoidsac2p
# add --adversary to drive the learned worst-case force too

Expected result¶

The ReachAvoidSAC2P policy holds a robust upright stance against the learned adversary and is the strongest safety policy on this stack. As a certificate it supplies the V(s) that the filter gauntlets use to protect the flat walker.

Known limitations¶

go2_locomote reuses the go2_stabilize twin as a fallback; that twin was trained at rest (cmd_vx = 0), so it is out-of-distribution when asked to hold a 1 m/s gait — a locomotion-specific fallback is needed for a strong filter result there.
The V ≥ 0 certificate from an older SAC twin is unsound (the entropy-bonus fix in safety_sb3 v0.4.0) — retrain on v0.4.0.

robot_safety_sandbox/tasks/go2_stabilize.py — the go2_stabilize and go2_locomote registrations.
robot_safety_sandbox/tasks/go2_velocity.py — the go2_walker_flat cumulative task policy.
configs/go2_stabilize_reachavoidsac2p.yaml, configs/go2_stabilize_reachavoidppo1p.yaml, configs/go2_walker_flat.yaml.