Launch a new mlagents-learn training run as a background process. Always overwrites previous results for the same run_id. To continue from a checkpoint, use resume_training instead.

By default, blocks until ready: in editor mode (no env_path), waits until mlagents-learn says 'Listening on port... press Play'. In batch mode (with env_path), waits until the executable connects.

Args: config_path: Path to the YAML training config file. run_id: Unique identifier for this training run. env_path: Path to a built Unity environment executable. Omit to use the Unity Editor. num_envs: Number of parallel Unity environment instances. no_graphics: Disable graphics rendering for faster training. torch_device: PyTorch device (e.g. 'cuda', 'cpu', 'cuda:0'). seed: Random seed (-1 for random). time_scale: Unity time scale multiplier. base_port: Base port for Unity communication (auto-assigned if omitted). wait: Block until ready (default true). Editor mode: waits for 'press Play'. Batch mode: waits for connection. wait_timeout: Max seconds to wait when wait=true.

Gracefully stop a running training run (sends SIGINT so the model is saved).

Args: run_id: The run to stop. timeout: Seconds to wait for graceful shutdown before force-killing.

Resume a previously stopped/completed training run.

If config_path is not provided, reads the saved configuration.yaml from the previous run. By default, blocks until ready (same as force_training).

Args: run_id: The run_id to resume. config_path: Config file path (auto-detected from previous run if omitted). num_envs: Number of parallel Unity environment instances. no_graphics: Disable graphics rendering. torch_device: PyTorch device. time_scale: Unity time scale multiplier. wait: Block until ready (default true). Editor mode: waits for 'press Play'. Batch mode: waits for connection. wait_timeout: Max seconds to wait when wait=true.

Kill orphaned mlagents-learn and Unity build processes that are not tracked by any active run. Use this when force_training fails due to leftover processes occupying ports, or after a crash.

Get detailed status of a training run including step progress, reward trend, and checkpoints.

Args: run_id: The run to query. last_n_rewards: Number of recent reward data points to include in the trend. last_n_checkpoints: Number of recent checkpoints to include.

Read TensorBoard scalar metrics from a training run.

Args: run_id: The run to query. behavior_name: Specific behavior to read (auto-detected if omitted). metric_keys: Specific metric keys to read (defaults to reward, losses, LR). last_n: Number of most recent data points per metric (default 20). Use -1 for all.

Get recent stdout/stderr output from an active training run.

Args: run_id: The run to query. last_n_lines: Number of most recent log lines to return.

List known training runs with their status.

Args: status_filter: Filter by status: 'running', 'completed', 'failed', or 'stopped'. last_n: Number of most recent runs to return (default 20). Use -1 for all.

Compare a specific metric across multiple training runs (useful for hyperparameter tuning).

Args: run_ids: List of run IDs to compare. metric_key: The TensorBoard metric key to compare (e.g. 'Environment/Cumulative Reward'). last_n: Number of most recent data points per run to include in the trend. Use -1 for all.

Locate .onnx model files and checkpoints for a training run.

Args: run_id: The run to query. behavior_name: Specific behavior (searches all if omitted). last_n: Number of most recent models/checkpoints to return. Use -1 for all.

Read a YAML training configuration file.

Args: config_path: Path to the config file (relative to config dir or absolute).

Deep-merge updates into an existing YAML config file. Only specified keys are changed.

Args: config_path: Path to the config file (relative to config dir or absolute). updates: Dictionary of updates to deep-merge into the config.

Block until the training run produces its first TensorBoard metric data point. Use this right after starting training to know when data starts flowing. Typically takes 1-2 minutes.

Args: run_id: The run to watch. timeout: Max seconds to wait. poll_interval: Seconds between checks.

Block until a training run finishes. Use this to chain runs automatically: start skill A, wait_for_completion, then start skill B. This WILL freeze the conversation until training ends or timeout. Default timeout is 4 hours.

Args: run_id: The run to wait for. timeout: Max seconds to wait (default 14400 = 4 hours). poll_interval: Seconds between internal status checks (default 60).

Check if training reached a target step count. Always returns instantly with current step, reward, and whether the target was reached.

Args: run_id: The run to check. target_step: The step count target.

Check if mean cumulative reward reached a target threshold. Always returns instantly with current reward, step, and whether the target was reached.

Args: run_id: The run to check. target_reward: The reward threshold.

Check if a training run has finished. Always returns instantly with current status, step, reward, and whether it completed.

Args: run_id: The run to check.

Check if new .onnx checkpoint files appeared on disk. Always returns instantly with list of all checkpoints and any new ones.

Args: run_id: The run to check. known_checkpoints: List of checkpoint paths already known (from a previous call). New files = current minus these. If omitted, returns all checkpoints.