---

name: baselines-training-and-evaluation description: "Use Habitat-Baselines training, evaluation, Hydra configs, trainers, checkpoints, and benchmark scripts safely." disable-model-invocation: true metadata: disco-role: operating license: MIT

Habitat-Baselines Training And Evaluation

Use this sub-skill when the task is about running, inspecting, or adapting Habitat-Baselines training and evaluation workflows rather than implementing Habitat core environments or new model code.

Read This First

• For CLI syntax, Hydra config groups, legacy command conversion, and safe config inspection, read references/cli-reference.md.
• For choosing RL/IL configs, training versus evaluation, checkpoint handling, TensorBoard/video outputs, and trainer APIs, read references/training-workflows.md.
• For Habitat 2.0/3.0 benchmark scripts, expensive runtime expectations, and safe adaptation patterns, read references/benchmarking.md.
• For PyTorch/CUDA, missing data/checkpoints, Hydra override errors, multiprocessing, DD-PPO, and video/TensorBoard failures, read references/troubleshooting.md.
• To inspect CLI availability or compose a baseline config without starting training, use scripts/baselines_cli_probe.py.

Use For

• Running habitat-baselines or python -m habitat_baselines.run with --config-name and Hydra overrides.
• Converting old --exp-config ... --run-type train/eval commands to current Hydra-style commands.
• Selecting PointNav, ObjectNav, ImageNav, InstanceImageNav, Rearrange, SocialNav, SocialRearrange, or EQA baseline configs.
• Switching among ppo, ddppo, ver, and IL trainers such as pacman, vqa, and eqa-cnn-pretrain.
• Resuming training, evaluating checkpoint files or checkpoint folders, and interpreting log, checkpoint, TensorBoard, video, and metric outputs.
• Deciding whether Habitat 2/3 benchmark scripts are safe to adapt or should be treated as reference-only.

Do Not Use For

• Creating Habitat tasks, datasets, sensors, measurements, or core Env APIs; use the tasks/datasets/envs sub-skill.
• Installation, package selection, and global config basics; use setup/configuration guidance first.
• HITL app or interactive policy integration; use the HITL sub-skill.
• Writing a new trainer, policy, registry entry, or observation transformer implementation; use extension-patterns.

Safe Default Workflow

1. Confirm that habitat-baselines --help or python -m habitat_baselines.run --help works before proposing a long run.
2. Use the probe script to list likely config names or compose the requested config with overrides.
3. If the user asks for actual training or evaluation, verify simulator support, dataset paths, checkpoint paths, GPU expectations, output directories, and whether the command is intended as a local smoke run or an expensive experiment.
4. Prefer short dry-run config composition and small override examples over running benchmark scripts or full training in an agent session.

Key Entry Points

• Console script: habitat-baselines, registered to habitat_baselines.run:main.
• Module CLI: python -m habitat_baselines.run.
• Programmatic config loader: habitat_baselines.config.default.get_config(config_path, overrides=None).
• Programmatic execution: habitat_baselines.run.execute_exp(config, "train") or execute_exp(config, "eval").
• Trainer registry: habitat_baselines.common.baseline_registry.baseline_registry.get_trainer(config.habitat_baselines.trainer_name).