API guide

This page is the landing map for the environment layer — tasks, margins, the registry, and the bridges to safety_sb3. The material that used to live here has been split by intent into concepts, reference, and user-guide pages; follow the links below. The algorithm layer (learners, backups, terminal_type) is documented in safety-stable-baselines docs/API.md.

Concepts

• Safety and reach-avoid margins — the g / l contract, the reach-avoid value, composing a margin_fn, and why avoid is not a reach-avoid instance.
• The MAP naming convention — Mode · Algorithm · Players, the algo_name formula, and why *PPO2P and *SAC2P are different algorithms.
• Episode termination — the end_criterion field.
• Safety-filter architecture — fallback · monitor · intervention.

Reference

• Task registry and Python API — TaskSpec fields, the registry functions, the two bridges, and configuration recipes.
• Code reference — the generated docstrings.
• Command-line reference — every flag of the trainers, the evaluator, and the viewer.

User guide

• Choose a workflow — the four primary workflows.
• Training — the trainer router, recipes, warm starts.
• Evaluation — the eval harness and its four axes.

Add a task or robot

• Tutorial: build the car-goal task — the full stack, from scratch.
• Extending and Porting a task.
• Installation and requirements.