Configuration

Config File

RoboSmith looks for robosmith.yaml or robosmith.yml in the command working directory unless you pass --config.

robosmith run --config robosmith.yaml --task "Walk forward"

Example:

llm:
  provider: anthropic
  model: anthropic/claude-sonnet-4-6
  fast_model: anthropic/claude-haiku-4-5-20251001
  temperature: 0.7
  max_retries: 3

reward_search:
  candidates_per_iteration: 4
  num_iterations: 3
  eval_timesteps: 50000
  eval_time_minutes: 2.0

training_backend: sb3
max_iterations: 3
skip_stages: []
scout_source: semantic_scholar
runs_dir: ./robosmith_runs
env_registry_path: ./my_envs.yaml

Loading Order

For robosmith run, the effective config comes from:

1. Built-in model defaults.
2. Auto-detected robosmith.yaml or robosmith.yml.
3. Explicit --config file.
4. CLI flags.
5. Runtime LLM provider detection from environment variables.

The --llm, --scout, --backend, --candidates, --skip, and task-related CLI flags override file values.

Environment Variables

RoboSmith loads .env.local and .env.

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GEMINI_API_KEY=AIza...
GROQ_API_KEY=gsk_...
OPENROUTER_API_KEY=sk-or-...
S2_API_KEY=...

S2_API_KEY is optional and only affects Semantic Scholar rate limits.

ForgeConfig

Field	Default	Meaning
llm	LLMConfig()	LLM provider, model, fast model, temperature, retries.
reward_search	RewardSearchConfig()	Reward candidate and evaluation budget.
runs_dir	./robosmith_runs	Base directory for artifacts and logs.
env_registry_path	None	Optional custom environment registry YAML.
max_iterations	3	Max outer loop iterations.
skip_stages	[]	Optional stages to skip: scout, intake, delivery.
scout_source	semantic_scholar	semantic_scholar, arxiv, or both.
verbose	True	Verbose behavior for configured callers.
dry_run	False	Parse and plan only.

LLMConfig

Field	Default	Meaning
provider	anthropic	Provider label.
model	claude-sonnet-4-20250514	Main model for code generation.
fast_model	claude-haiku-4-5-20251001	Fast model for routing and parsing.
temperature	0.7	Sampling temperature.
max_retries	3	Retries for LLM calls.

RewardSearchConfig

Field	Default	Meaning
candidates_per_iteration	4	Number of reward candidates generated per generation.
num_iterations	3	Reward evolution iterations.
eval_timesteps	50000	Short evaluation steps per candidate.
eval_time_minutes	2.0	Max candidate evaluation time.

TaskSpec

TaskSpec can be built directly from Python or produced by intake.

from robosmith.config import TaskSpec

spec = TaskSpec(
    task_description="A Franka arm picks up a red cube",
    robot_type="arm",
    robot_model="franka",
    environment_type="tabletop",
    algorithm="auto",
    time_budget_minutes=60,
    num_envs=1024,
)

Important fields:

Field	Meaning
task_description	Natural language desired behavior.
raw_input	Preserved original user input.
robot_type, robot_model	Morphology and optional exact robot.
environment_type, environment_id	Environment class and optional forced registry ID.
success_criteria	List of SuccessCriterion.
safety_constraints	List of SafetyConstraint.
algorithm	ppo, sac, td3, or auto.
time_budget_minutes, num_envs	Training budget hints.
use_world_model	Future world-model pretraining switch.
push_to_hub	Optional HuggingFace repo ID.