Robot Safety Sandbox

Parallelized mjlab environments for nominal-policy training, safety-policy synthesis, and safety-filter evaluation.

Massively-parallel mjlab benchmark environments for safety_sb3 (safety-stable-baselines): reach-avoid / avoid-only × single-agent / adversarial (ISAACS), on GPU end-to-end — plus a filters/ library with the three deployment styles (value shielding, R-CBF/Q-CBF projection, rollout shielding).

Renamed from safe_mjlab_zoo (a deprecated import alias remains for one transition cycle).

from robot_safety_sandbox import make_tensor, list_tasks
from safety_sb3 import ReachAvoidPPO

env = make_tensor("go2_gap_chain", num_envs=2048)      # ~50k steps/s on 12 GB
model = ReachAvoidPPO("MlpPolicy", env, normalize_obs=True, adaptive_lr=True,
                      ent_coef=1e-4, n_steps=48, batch_size=24576,
                      policy_kwargs=dict(log_std_init=-1.204))
model.learn(2_000_000_000)

The contract (what every task guarantees)

channel	meaning
reward	g(s) — physical safety margin. Never normalize or reshape it.
l_x	l(s) — target margin (zeros for avoid-only tasks)
dones / timeouts	mjlab auto-resets; timeouts are never value-bootstrapped
metrics()	curriculum levels + task metrics, forwarded to the logger every rollout

A task = cfg_builder(play) -> ManagerBasedRlEnvCfg (spawn events, curricula, terrain — plain mjlab, algorithm-agnostic) + margin_fn(env) -> (g, l) (compose from margins.py). Register a TaskSpec and both bridges (make_tensor for PPO learners, make_numpy for the SAC family) work.

Tasks

task	objective	learner	warm-starts from
go2_stabilize / go2_locomote	stand / track a command vs adversarial force (the original task; simplest zoo entry)	ReachAvoidPPO	—
digit_stabilize	humanoid stand vs adversarial torso force (Digit analog of go2_stabilize)	ReachAvoidPPO	—
go2_gap_landing	soft-land from mid-air over a gap	SafetyPPO	—
go2_gap_crossing	reverse curriculum: landing → launch	SafetyPPO	landing
go2_gap_chain	takeover momentum → safe rest (brake/jump)	ReachAvoidPPO	crossing
go2_gap_chain_isaacs	chain + worst-case force adversary	IsaacsPPO	chain
go2_crawl / _isaacs	duck under a low bar or stop	ReachAvoidPPO	—

Task structure varies: go2_stabilize needs no curriculum or staging at all, while the gap family only forms its jump through staged warm-starts (TaskSpec.warmstart_from) at real scale (~2B env-steps for the chain). See PORTING.md for which machinery your task actually needs.

Training recipe that works (hard-won)

normalize_obs=True (obs only — reward normalization is refused by safety_sb3), ent_coef=1e-4, log_std_init=ln(0.3), adaptive_lr=True (desired_kl=0.01, lr 5e-4), n_steps=48. Watch the env/Curriculum/* logger keys — a stalled curriculum looks exactly like converged training in the reward curve.

Porting a new task

1. Write the mjlab env cfg: terrain, spawn events (takeover-momentum or staged spawns), reverse curricula. Study go2_gap — especially how the landing → crossing → chain pipeline seeds rare-win skills.
2. Compose margin_fn from margins.py (or add new terms there).
3. register(TaskSpec(...)) in tasks/<your_task>.py.
4. Train with examples/train.py --task <id>; verify curricula CLIMB in wandb.

Extending

docs/EXTENDING.md walks the four extension axes with worked examples from the shipped tasks: margin functions, sensors/observations, terrains (heightfields, walls, gaps, obstacles), and contacts — plus the new-robot checklist (go2 and digit are the two reference layouts).

Repo layout / status

SELF-CONTAINED: env cfgs, terrains, robot assets (Go2, Digit), and the handover dataset are native under robot_safety_sandbox/envs/ + data/. safety_sb3 is a pinned pip dependency (safety-stable-baselines v0.1.0); mjlab is a peer dep with its own pinned sim stack (INSTALL.md).