orbit-robotics 0.3.0

Data engineering copilot for robot imitation learning datasets

ORBIT

Orbit analyzes your robot training data and tells you what's wrong before you waste 8 hours training.

pip install orbit-robotics
orbit analyze lerobot/your-dataset

The Workflow

1. Analyze your dataset

orbit analyze lerobot/xarm_lift_medium

Dataset Readiness: C (score: 65/100)
Usable but has problems — run orbit clean first

  ✓ High consistency (1.00)
  ✓ Sufficient episodes (800) for diffusion_policy
  ✓ Good policy fit (1.00)
  ✗ 4 joints clipping (>10% of frames)
  ✗ High action divergence (0.46) — demos contradict each other in similar states

YOUR DATA AT A GLANCE
  Episodes:       800     (top 25%)
  Coverage:       0.84  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░
  Signal Health:  0.00  ░░░░░░░░░░░░░░░░░░░░░░░░░

2. Clean bad episodes

orbit clean lerobot/xarm_lift_medium --dry-run

┌────────────────────┬──────────┐
│ Original episodes  │ 800      │
│ Episodes to remove │ 0 (0%)   │
│ Episodes to keep   │ 800      │
│ Dead joints        │ 0 of 4   │
│ Quality            │ 71 → 71  │
└────────────────────┴──────────┘

When bad episodes are found, orbit clean removes them using the LeRobot API and tells you exactly what was wrong with each one.

3. Get a training command

orbit suggest lerobot/xarm_lift_medium

Recommended: Diffusion Policy (fit: 1.00)
┌─────────────┬──────────────────────┐
│ Episodes    │ 800                  │
│ Cameras     │ 1 x 480p             │
│ Action dims │ 4                    │
│ FPS         │ 15                   │
│ Alternative │ SmolVLA (fit: 1.00)  │
└─────────────┴──────────────────────┘

lerobot-train \
  --dataset.repo_id=lerobot/xarm_lift_medium \
  --policy.type=diffusion_policy \
  --batch_size=32 \
  --steps=500000 \
  --policy.horizon=4 \
  ...

What Orbit Catches

• Dead servos that stopped moving during data collection — wasting model capacity learning zero outputs
• Aborted or corrupted episodes that are too short, too long, or have no meaningful motion
• Clipping joints hitting their position limits, which creates discontinuous action targets
• Inconsistent demonstrations where you did different things in similar states, confusing behavior cloning
• Wrong policy for your data — ACT needs consistent demos, Diffusion Policy handles multimodal strategies, SmolVLA needs fewer episodes but wants cameras
• Not enough data for your chosen policy, with specific targets: ACT wants 50+, Diffusion Policy wants 100+, BC wants 200+
• Timing problems — frame drops, FPS jitter, and state-action lag from slow teleoperation setups
• Low workspace coverage — demonstrations that only cover a narrow region of the task space

More Commands

# Plan a data collection session
orbit plan "pick up cups" --robot so100 --policy act

# Diagnose training failures from logs
orbit debug outputs/train/my-run

# Browse 82 published benchmarks
orbit benchmark --task pick_and_place --min-success 0.7

# Track progress against your collection plan
orbit track lerobot/my-dataset

# JSON output for scripting
orbit analyze lerobot/my-dataset --json

# Deep AI analysis (requires API key)
orbit analyze lerobot/my-dataset --deep

Policy Support

Policy	Flag	Best for
ACT	--policy act	Consistent, high-res demos (50+ episodes)
Diffusion Policy	--policy diffusion_policy	Multimodal strategies (100+ episodes)
SmolVLA	--policy smolvla	Vision-language tasks, fewer episodes needed
DP3	--policy dp3	3D point cloud observations
BC / BC-RNN	--policy bc	Large datasets (200+ episodes)

--policy auto (default) recommends the best policy for your data.

Install

pip install orbit-robotics          # Core analysis (no GPU needed)
pip install orbit-robotics[vision]  # + SigLIP embedding analysis
pip install orbit-robotics[vlm]     # + Gemini VLM task analysis
pip install orbit-robotics[all]     # Everything

License

MIT — see LICENSE for details.