Training Backends

One Trainer Interface

RoboSmith uses a universal trainer interface so the pipeline can call SB3, CleanRL, rl_games, imitation learning, offline RL, or future backends through the same contract.

from robosmith.trainers.registry import TrainerRegistry

registry = TrainerRegistry()
trainer = registry.get_trainer(algorithm="ppo", backend="sb3")

Built-In Backends

Backend	Class	Paradigm	Algorithms
sb3	SB3Trainer	rl	PPO, SAC, TD3, A2C, DQN
cleanrl	CleanRLTrainer	rl	PPO
rl_games	RLGamesTrainer	rl	PPO
il_trainer	ILTrainer	il	BC, DAgger
offline_rl_trainer	OfflineRLTrainer	offline_rl	TD3+BC, CQL, IQL

Check local availability:

robosmith trainers

TrainingConfig

from robosmith.trainers.base import TrainingConfig, LearningParadigm

config = TrainingConfig(
    task_description="A quadruped walks forward",
    algorithm="ppo",
    paradigm=LearningParadigm.REINFORCEMENT_LEARNING,
    env_id="Ant-v5",
    total_timesteps=50_000,
    time_limit_seconds=300,
    seed=42,
    device="auto",
    extra={"n_steps": 2048},
)

Fields are intentionally broad enough for multiple learning paradigms:

Field	Meaning
env_id, env_entry	Environment identity and registry entry.
reward_fn	Custom reward function for RL.
demo_paths, num_demos	Demonstration data for imitation learning.
dataset_path	Static dataset for offline RL or VLA-style methods.
total_timesteps, total_epochs	Budget by training style.
artifacts_dir	Where checkpoints and logs should be written.
extra	Backend-specific options.

TrainingResult

Every trainer returns TrainingResult.

result = trainer.train(config)

if result.success:
    print(result.model_path)
    print(result.final_mean_reward)
else:
    print(result.error)

Important fields:

Field	Meaning
model_path	Path to the saved model or checkpoint.
algorithm	Algorithm actually used.
paradigm	Learning paradigm.
total_timesteps	Completed environment steps.
training_time_seconds	Wall-clock training time.
final_mean_reward, final_std_reward	Final reward metrics.
converged	Backend-specific convergence signal.
metrics_history	Time series or backend metrics.
error	Error string when training failed.

Policy Selection

select_policy_approach() picks a learning paradigm, algorithm, and backend from task context.

from robosmith.trainers.selector import select_policy_approach

approach = select_policy_approach(
    task_description="Pick and place a cube",
    env_entry=entry,
    available_backends=["sb3", "cleanrl"],
)
print(approach.algorithm)
print(approach.reason)

Selection rules:

Signal	Typical choice
Many demonstrations	Imitation learning with BC.
Large static dataset	Offline RL with IQL.
Discrete action space	PPO.
Classic control	PPO.
Locomotion	PPO, rl_games if GPU and available.
General manipulation	SAC.
Dexterous manipulation	TD3, with SAC as an alternative.

Forcing A Backend

robosmith run --task "..." --backend sb3
robosmith run --task "..." --backend cleanrl
robosmith run --task "..." --algo sac

When a backend is forced, TrainerRegistry.get_trainer() verifies it exists and that its dependencies are installed.