Skip to content

Environments catalog

The mjlab / GPU robot benchmark environments that ship with robot_safety_sandbox — reach-avoid, avoid-only, and dense-reward, single-agent and adversarial, end to end on GPU. Each is a TaskSpec in the registry; make_tensor("<id>") builds it for a safety_sb3 learner.

Learner names follow the MAP — the names shown are the on-policy (PPO) family; swap PPO→SAC for --family off_policy.

Maturity labels

list_tasks() deliberately exposes internal and experimental variants alongside the primary public tasks. Use these labels to tell them apart: Tutorial · Stable (reproduced, recommended) · Research (a shipped result; needs staged training / real scale) · Experimental (shelved or in progress; may not reproduce) · Requires custom mjlab fork (imports on stock mjlab, needs a fork to simulate).

  • Car goal · Tutorial

    car goal

    A differential-drive car reaches a goal while avoiding obstacles. The new-user walkthrough, built from scratch. ReachAvoidPPO1P

  • Go2 stabilize / locomotion · Stable

    A quadruped stays upright (and tracks a command) against a worst-case force. The simplest and most-reproduced safety benchmark. ReachAvoid* / Safety*

  • Go2 gap-jumping · Research

    go2 gap

    A quadruped brakes or commits to a leap over a pit — the flagship reach-avoid pipeline. SafetyPPO1P · ReachAvoidPPO1P · ReachAvoidPPO2P

  • Digit stabilize · Requires custom mjlab fork

    digit

    A humanoid stays upright against a worst-case torso force — two-player avoid. SafetyPPO1P · SafetyPPO2P

Also documented: Go2 crawl (Experimental — the low-bar / closing-gate crawl thread).

Primary tasks

Task ID Robot Mode Objective Adversary Maturity
car_goal diff-drive car reach-avoid drive to a goal disk, avoid obstacle cylinders — Tutorial
go2_stabilize Go2 reach-avoid stand / stay upright vs adversarial base force ✔ Stable
go2_locomote Go2 reach-avoid track a velocity command vs adversarial force ✔ Research
go2_walker_flat Go2 cumulative dense-reward flat walker — the task policy filters wrap — Stable
go2_walker_filtered Go2 cumulative task policy trained inside a filter (PORL showcase) eval only¹ Research

The Adversary column: ✔ means the task supports adversarial training (--adversary resolves a two-player *2P learner). ¹go2_walker_filtered is cumulative, so it has no two-player learner — the trainer refuses --adversary. Its supports_adversary=True only exposes an eval-time disturbance channel (a live --dstb / --dstb-scale attack in examples/eval.py), not adversarial training.

Go2 gap-jumping family (Research)

The jump forms through staged warm-starts; the split-test twins isolate the reach-avoid-vs-avoid contrast. See Go2 gap-jumping.

Task ID Mode Objective Learner (1P)
go2_gap_landing safety soft-land from mid-air over the gap SafetyPPO1P
go2_gap_crossing safety reverse curriculum: landing → launch SafetyPPO1P
go2_gap_chain reach-avoid arrival momentum → brake-or-jump → safe rest ReachAvoidPPO1P
go2_gap_chain_isaacs reach-avoid chain + worst-case base-force adversary ReachAvoidPPO1P (--adversary → 2P)
go2_gap_brake_or_jump_ra (_w20, _w30) reach-avoid split-test reach-avoid twin ReachAvoidPPO1P
go2_gap_brake_or_jump_avoid (_w20, _w30) safety split-test avoid twin SafetyPPO1P

Digit (Requires custom mjlab fork to simulate)

Assets and cfgs are vendored and import cleanly on stock mjlab; stepping the sim needs the lab's mjlab entity-patch fork. See Digit stabilize.

Task ID Mode Objective
digit_stabilize reach-avoid stand / track vs adversarial torso force
digit_stabilize_avoid · digit_stabilize_stay safety stay upright forever
digit_box_stabilize_avoid · digit_box_stabilize_stay safety as above + keep a box balanced

All Digit tasks declare supports_adversary (two-player avoid / reach-avoid).

Experimental tasks (folded threads)

The crawl / low-bar / tunnel line is shelved (a model-free reach-avoid executor hit a ~0.3 success ceiling on sustained crawling); the environments remain registered for reference and may not reproduce a clean result. These appear in list_tasks() but are not recommended entry points.

Family Task IDs Mode(s)
Crawl go2_crawl, go2_crawl_duck, go2_crawl_gate_ra / _gate_avoid, go2_crawl_twin_ra / _twin_avoid, go2_crawl_isaacs, go2_crawl_locomote, go2_crawl_walk reach-avoid / safety / cumulative
Low bar go2_low_bar_ra / _avoid (+ _h70 / _h80 / _h90), go2_low_bar_gate_ra / _gate_avoid, go2_low_bar_bridge reach-avoid / safety / cumulative
Tunnel go2_tunnel_ra / go2_tunnel_avoid reach-avoid / safety
Rendering-only² go2_crawl_duck_video, go2_crawl_walk_video reach-avoid, cumulative

²"Rendering-only" is a use/maturity classification (these variants exist to capture rollout GIFs), not a mode — each still has a real mode: go2_crawl_duck_video is reach-avoid, go2_crawl_walk_video is cumulative.

The algorithm layer's CPU reference environments

For the small CPU reference environments (bicycle, pendulum, …) that ship with the algorithm layer, see the safety-stable-baselines environment showreel.

Page template

The Stable and Research pages (Go2 stabilize, Go2 gap-jumping) follow the full template: a compact operational summary (status / requirements, task IDs, success/failure, spaces, margins, config, budget, train / evaluate / visualize, expected result, limitations, source files) followed by the research narrative. Pages for tasks that cannot report reproduced results on stock mjlab carry a reduced template — Digit is gated behind a custom mjlab fork (it cannot be simulated on stock mjlab), and Go2 crawl is a shelved/experimental thread with no clean result; both document Tasks, Margins, and Run only. GIFs are captured from evaluation rollouts; see each page for the examples/eval.py command.