orbit-robotics 0.6.0

Data engineering copilot for robot imitation learning datasets

ORBIT — Data Quality Analysis for Robot Policy Training

Predict whether your robot learning dataset will actually train successfully — before burning GPU hours.

ORBIT analyzes your demonstration data for the issues that cause training failures: dead joints, action divergence, inconsistent demonstrations, poor workspace coverage, and more. Get a quality grade (A-F) and calibrated success prediction based on 82+ ground-truth training outcomes.

Quick Start

pip install orbit-robotics
orbit analyze lerobot/pusht

No GPU required. No API keys required. Works out of the box.

What It Checks

• Quality grade (A-F) calibrated against real training outcomes
• Success prediction — probability your data will train a working policy
• Dead joint detection — finds servos that aren't moving
• Action divergence — detects contradictory demonstrations
• Episode consistency — flags recording issues and length outliers
• Policy fit — rates compatibility with ACT, Diffusion Policy, SmolVLA, OpenVLA
• Workspace coverage — checks if demonstrations cover the task space
• Community comparison — benchmarks against other public datasets

Example Output

ORBIT Analysis: lerobot/pusht
206 episodes - DIFFUSION_POLICY

  Grade: A (98/100) — Ready to train — expect strong results
  Similar datasets trained at: 63%, 84%, 91%, 95%, 100% (5 nearest matches)

  1 issue found:
    ! Jerk varies across episodes (CV=0.57)

  Run with --proxy for training signal, --detail for full diagnostics

Local AI — Free, Private, Zero Config

ORBIT v0.6 adds local AI support via Ollama. Deep analysis, AI grading, and the orbit assist chatbot all run locally with no API keys and no data leaving your machine.

# Install Ollama (one time)
curl -fsSL https://ollama.com/install.sh | sh
ollama pull gemma4

# Now AI features just work — ORBIT auto-detects Ollama
orbit analyze lerobot/your-dataset --deep    # AI-powered deep analysis
orbit analyze lerobot/your-dataset --ai      # AI quality judge
orbit assist                                 # Interactive AI troubleshooter

ORBIT auto-detects the best available provider: Ollama (local) > Gemini (cloud) > OpenAI (cloud). No configuration needed — it just picks whatever's available.

Check your AI setup anytime:

orbit setup-ai

Provider Options

Provider	Setup	Cost	Data Privacy
Ollama (recommended)	ollama pull gemma4	Free	Data stays local
Gemini	export GOOGLE_API_KEY=...	~$0.001/analysis	Sent to Google
OpenAI	export OPENAI_API_KEY=...	~$0.01/analysis	Sent to OpenAI

Override the auto-detected provider:

orbit auth set ai-provider ollama      # Force Ollama
orbit auth set ai-model gemma4         # Pick a specific model

Commands

Command	What it does
orbit analyze <dataset>	Full quality analysis with grade and predictions
orbit suggest <dataset>	Training command with tuned hyperparameters
orbit clean <dataset>	Remove bad episodes automatically
orbit fix <dataset>	Analyze, clean, and suggest in one shot
orbit gate <dataset>	CI/CD quality gate — pass/fail for pipelines
orbit compare <a> <b>	Side-by-side dataset comparison
orbit benchmark <task>	Compare against published training benchmarks
orbit assist	AI troubleshooter for data and training issues
orbit doctor	Check environment health and AI providers
orbit setup-ai	Check and configure AI providers
orbit badge <dataset>	Generate a shields.io quality badge
orbit explore	Browse and discover LeRobot datasets on HuggingFace

Common Options

orbit analyze lerobot/my-dataset --policy act       # Check fit for specific policy
orbit analyze lerobot/my-dataset --deep              # AI-powered deep analysis
orbit analyze lerobot/my-dataset --ai                # AI quality judge for A grades
orbit analyze lerobot/my-dataset --json              # Machine-readable output
orbit analyze lerobot/my-dataset --full              # All episodes (no sampling)
orbit analyze lerobot/my-dataset --detail            # Full diagnostic report
orbit -q analyze lerobot/my-dataset                  # Quiet mode (grade only)
orbit analyze ./local-data/ --format hdf5            # Local HDF5 files

CI/CD Integration

# Fail the pipeline if data quality is below B
orbit gate lerobot/my-dataset --min-grade B --policy act

# JSON output for scripts
orbit analyze lerobot/my-dataset --json | jq '.readiness.grade'

Supported Formats

Format	Source
LeRobot (Hub)	HuggingFace datasets (lerobot/...)
LeRobot (local)	Local LeRobot directories
HDF5	RoboMimic, robosuite, custom .hdf5 files
RLDS	TFRecord-based datasets (pip install orbit-robotics[rlds])
ROS bags	.bag and .mcap files (pip install orbit-robotics[rosbag])

Understanding Grades

Grade	Score	Meaning
A	85-100	Ready to train — expect strong results
B	72-84	Good data — minor issues, should train well
C	58-71	Usable but has problems — clean first
D	40-57	Significant issues — collect more or better data
F	0-39	Critical problems — fix before training

Grades are calibrated against 82 real datasets with known training outcomes.

What's New in v0.6.0

• Local AI via Ollama — --deep, --ai, and orbit assist work locally with Gemma 4, no API keys needed
• Multi-provider AI — auto-detects Ollama > Gemini > OpenAI, configurable via orbit auth set ai-provider
• orbit setup-ai — check and configure AI providers in one command
• Grading accuracy overhaul — ground truth task difficulty adjustment, BC/BC-RNN divergence penalties, calibrated summaries for hard tasks
• Similar dataset display — shows nearest benchmark success rates instead of wide confidence intervals
• Sampling warnings — warns when analyzing <20 episodes that the grade may be unreliable
• orbit doctor — now checks Ollama, Gemini, and OpenAI availability
• Ollama backend for orbit assist — chat with local models about your data

Requirements

• Python 3.10+
• No GPU needed
• No API keys needed for core analysis

License

MIT