Integration Tooling

Why It Exists

Existing robotics artifacts rarely share the same interface. A policy may expect different camera names than a dataset provides. A dataset may store actions with one extra gripper dimension. An environment may expose a dict observation while a trainer expects flat vectors.

The integration tooling makes those interfaces visible before generating code.

Inspect

robosmith inspect turns artifacts into structured Pydantic result models. The default output is human-readable Rich tables; --json returns machine-readable JSON.

robosmith inspect dataset lerobot/aloha_mobile_cabinet
robosmith inspect dataset lerobot/aloha_mobile_cabinet --schema --quality --json
robosmith inspect env Ant-v5 --obs-docs --sample
robosmith inspect policy lerobot/smolvla_base
robosmith inspect robot path/to/robot.urdf

Inspection result models:

Artifact	Result model
Dataset	DatasetInspectResult
Environment	EnvInspectResult
Policy	PolicyInspectResult
Robot description	RobotInspectResult
Compatibility report	CompatReport

Compatibility

robosmith inspect compat POLICY TARGET
robosmith inspect compat POLICY TARGET --fix
robosmith inspect compat POLICY DATASET ENV --json

Compatibility reports group issues by severity:

Severity	Meaning
critical	The artifacts cannot be used together without a fix.
warning	The artifacts may work, but behavior or performance is risky.
info	Context that may matter during integration.

The --fix flag uses the template generator to print a wrapper when the report is not compatible.

Generate

robosmith gen wrapper generates Python adapter code between a policy and a target dataset or environment.

# Uses the LLM generator by default.
robosmith gen wrapper lerobot/smolvla_base lerobot/aloha_mobile_cabinet

# Template path, no API key.
robosmith gen wrapper lerobot/smolvla_base lerobot/aloha_mobile_cabinet --no-llm

# Write to a file.
robosmith gen wrapper lerobot/smolvla_base lerobot/aloha_mobile_cabinet -o adapter.py

Generated wrappers are expected to address camera remapping, image resizing, action dimension adaptation, and normalization hooks where those mismatches are visible in the inspection results.

Diagnostics

robosmith diag reads rollouts and compares trajectory sets.

robosmith diag trajectory path/to/rollout.hdf5
robosmith diag trajectory lerobot/aloha_mobile_cabinet --json
robosmith diag compare rollout_a.hdf5 rollout_b.hdf5

Trajectory diagnostics report:

Field	Meaning
success_rate	Fraction of successful episodes when success labels are available.
episode_length_mean	Average episode length.
action_stats	Per-dimension mean, std, min, max, and clipping rate.
failure_clusters	Grouped failure descriptions when available.

Auto Integrate

robosmith auto integrate chains inspection, compatibility checking, wrapper generation, and validation into a single LangGraph workflow.

robosmith auto integrate \
  lerobot/smolvla_base \
  lerobot/aloha_mobile_cabinet \
  --verbose \
  -o adapter.py

The graph currently runs:

1. Inspect policy.
2. Inspect target.
3. Check compatibility.
4. Generate wrapper if needed.
5. Validate enough structure to report success or failure.

Use --json when another tool should consume the final status, warnings, and output file list.

Python Entry Points

from robosmith.inspect.dispatch import inspect_dataset, inspect_env, inspect_policy
from robosmith.inspect.compat import check_compatibility
from robosmith.generators.gen_wrapper import generate_wrapper
from robosmith.diagnostics.trajectory_analyzer import analyze_trajectory

dataset = inspect_dataset("lerobot/aloha_mobile_cabinet")
policy = inspect_policy("lerobot/smolvla_base")
report = check_compatibility(policy.model_id, dataset.repo_id)
code = generate_wrapper(policy.model_id, dataset.repo_id, use_llm=False)
diag = analyze_trajectory("path/to/rollout.hdf5")