---

name: smarts description: "Use SMARTS 2.0.1 for multi-agent autonomous-driving simulation, Gymnasium environments, scenario generation, sensor and Envision workflows, agent-zoo/RL integration, and the scl command-line tools." disable-model-invocation: true metadata: disco-role: operating license: MIT

SMARTS

Use this repo skill when the task involves SMARTS (Scalable Multi-Agent Reinforcement Learning Training School), autonomous-driving simulation, scenario maps and traffic, multi-agent Gymnasium environments, agent policies, road/sensor observations, Envision replay, or the scl CLI.

Operating contract

1. Identify the workflow before acting: environment lifecycle, scenario authoring, sensors/visualization, policy/RL/agent-zoo packaging, or CLI and external integrations.
2. Read the owning sub-skill and its linked references before writing code or commands. Keep paths to scenarios, models, records, and outputs explicit; do not assume a SMARTS source checkout.
3. Start with read-only import, signature, and CLI-help checks. Build or clean scenario artifacts, launch services, install optional stacks, and run training/benchmarks only after the user explicitly needs that operation and its inputs are validated.
4. Treat the core baseline as CPU-verified. Camera/Panda3D, Envision service, SUMO/TraCI, Ray/RLlib, Torch/TensorFlow, ROS, Waymo, Argoverse, Visdom, and downloaded datasets are optional capability boundaries; never infer their runtime success from a base import.
5. Keep a bounded run and deterministic seed for smoke tests. Close every environment/service/client that the task starts.

Install and verify

For the core Python API:

python -m pip install "smarts"
python -c "import smarts; print('SMARTS import OK')"

Add only the extras required by the workflow:

python -m pip install "smarts[gymnasium]"       # Gymnasium environments
python -m pip install "smarts[camera-obs]"      # camera/grid/Panda3D paths
python -m pip install "smarts[envision]"        # Envision client/server path
python -m pip install "smarts[sumo]"             # SUMO-backed traffic/maps
python -m pip install "smarts[ray,rllib]"        # optional Ray/RLlib stack

The exact extra names and compatible versions can change; inspect installed metadata and run the shared diagnostic before relying on an optional branch: scripts/check_smarts_install.py. The prepared baseline verified package smarts 2.0.1, core imports, Gymnasium, OpenDRIVE, Envision/Panda3D imports, and scl help on Python 3.11. It did not verify external SUMO/ROS/Waymo/Argoverse services or RL training stacks.

Route the task

User intent	Read first	Typical deliverable
Create/reset/step a single or multi-agent simulator, configure interfaces, observations, actions, rewards, or controllers	simulation-environments	Bounded HiWayEnvV1/parallel environment loop and policy contract
Define maps, traffic, missions, social actors, bubbles, friction, or replay histories; generate or validate scenario assets	scenario-studio	smarts.sstudio DSL, deterministic generation, layout validation
Add lidar, grid/RGB/occlusion/signal sensors, camera rendering, Envision recording, or replay	sensors-visualization	Sensor interface, renderer/service probe, record inspection
Implement a policy, AgentSpec, locator, social agent, bubble, RLlib adapter, or benchmark inference package	rl-agent-zoo	Importable/versioned agent package and optional RL integration plan
Use scl, diagnostics, scenario build/clean, benchmarks, zoo commands, SUMO/TraCI, Waymo, OpenDRIVE, or ROS boundaries	cli-integrations	Safe command plan, prerequisite probe, and recovery steps

Cross-route handoffs are deliberate: scenario-studio owns authoring while cli-integrations owns scl scenario build; simulation-environments owns the reset/step contract while sensors-visualization owns rendered observation prerequisites; rl-agent-zoo owns policies while CLI owns zoo/benchmark commands.

Core safety rules

• A scenario must already have the generated map/traffic artifacts before a core environment smoke. Use scenario-studio and then the CLI route to build it; do not silently substitute a different map.
• In multi-agent mode, observations and actions contain only currently active agents. An empty active-agent dictionary can be a valid intermediate state; keep stepping with {} until the global done flag or the bound is reached.
• Match policy outputs to env.action_space[agent_id]; do not guess action tuples from an interface name. Use space.contains(action) when debugging.
• Use headless=True for CPU checks. Image observations can require the camera extra and a display/offscreen backend, and can dominate step time.
• scl --help, scl scenario --help, and scl run --help are safe first checks. --clean, --auto-install, server startup, benchmark execution, and external-service commands have side effects or trust boundaries.

Read references/troubleshooting.md for cross-cutting installation, optional-extra, path/configuration, port, and lifecycle failures. Read references/repo-provenance.md before deciding whether this graph is stale for another SMARTS checkout.