Robotics Observability Platform
Open-source · Early Access · MatchArm-compatible

Debug robots
like software.

ROBOLOGR is a flight recorder + dataset quality toolkit for robotic systems. Align multi-camera video, actions, joint states, and hardware health on a single timeline. Know what happened, why it failed, and whether the data is worth training on.

View on GitHub See Modules ↓
robologr viewer — matcharm_0217.episode
matcharm_0217
47.2s · 3 cameras · 6 joints
QS 86 / 100
cam_front
joint states
torque / current
events
86/100
Quality Score
3
Anomalies
2.3ms
Avg Latency
⚠
t=19.4s — Joint 3 torque spike detected. Command vs actual lag: +38ms. Possible motor saturation or collision precursor.
6 signals
synchronized on one timeline
∞ episodes
local-first, no cloud lock-in
100pt QS
per-frame quality scoring
0 rewrites
export-ready training manifests
Modules

Everything your robot
data pipeline needs.

Six focused modules that go from raw sensor dump to training-ready, failure-annotated dataset.

🎥
Robot Observability
Episode bundles: multi-cam video, actions, state, events, and derived signals all aligned on a single interactive timeline. Scrub, zoom, and inspect any moment.
episode · timeline · scrubber
📊
Dataset Quality Scoring
Blur, exposure, occlusion, dropped frames, timestamp drift, background clutter — scored per-frame and per-episode using VLMs. Know before you train.
VLM · per-frame · QS 0–100
🏷️
Intent + Failure Labels
Define what "good" and "bad" looks like for your task. Tag intent and failure taxonomy with evidence — actual frames and signal clips, not just metadata.
labeling · taxonomy · evidence
💀
Failure-First Training
Build training splits with guaranteed failure coverage. Generate precursor reports so your model learns from mistakes — and never repeats them.
splits · precursors · coverage
🔧
Hardware Anomaly Detection
Current drift, overheating, comm drops, command-vs-actual lag. Separate hardware failures from policy failures so you fix the right thing.
anomaly · motors · health
📦
Training Exports
Export manifests for any training pipeline — no full dataset rewrites. Shareable run reports with embedded clips, signal charts, and quality breakdowns.
manifests · reports · adapters
How it works

Local-first.
No cloud required.

Runs next to your data. Indexes large datasets, streams only previews and metadata to the UI. Your data never leaves your machine.

01

Init your data directory

Point ROBOLOGR at any folder of episodes — ROS2 bags, custom formats, or raw video + CSV. It creates a local index.

02

Run quality + anomaly checks

Process runs per-episode, scoring visual quality, detecting signal anomalies, and flagging timestamp drift.

03

Inspect in the viewer

Open the browser UI, scrub timelines, review failure clips, edit labels, and annotate intent tags.

04

Export training manifests

Generate split files with failure guarantees, quality thresholds, and precursor reports. Plug into your pipeline.

terminal ● live
$ robologr init /data/matcharm
→ indexed 148 episodes
$ robologr process --quality --anomaly
→ QS: avg 81 · 12 flagged
→ anomalies: 7 torque, 2 comm
$ robologr serve
→ UI ready: http://localhost:7771
$ robologr export --split train --failure-min 0.3
→ manifest.json written · 94 episodes
Failure Intel

Train on what went
wrong.

Most datasets hide failures. ROBOLOGR surfaces them, labels them, and makes sure your model learns from every mistake before it deploys.

Joint 3 — Torque Anomaly
FAILURE
t=18.90torque nominal: 0.42 Nm
t=19.22⚠ torque rising: 1.1 Nm
t=19.44✖ spike detected: 3.8 Nm (threshold: 2.0)
t=19.52✖ cmd–actual lag: +38ms
t=19.80⚠ recovery: 1.2 Nm
t=20.10nominal restored
cam_wrist — Visual Quality
TRAINABLE
blur0.3% frames flagged (threshold: 5%)
exposurenominal — Δ EV < 0.4
occlusion0 frames > 30% blocked
timestampsdrift: 0.8ms max — PASS
clutter⚠ background complexity: high
verdict✓ TRAINING-WORTHY · QS 91
Open-Source First

Built in the open.
From day one.

Start with a stable episode format and viewer, then grow with adapters, QA, labeling, and anomaly modules. Community-driven, MatchArm-friendly.

✓
Episode format + viewer
Stable schema · multi-stream scrubber · timeline UI
shipped
▶
Dataset quality scoring
VLM-powered · per-frame + per-episode QS
in progress
▶
Hardware anomaly detection
Motor health · comm drops · lag analysis
in progress
○
ROS2 + custom recorder adapters
Ingest from any robotics stack
planned
○
Failure labeling + intent taxonomy
Evidence-linked labels · failure-first splits
planned
○
Training manifest exports
Pipeline-ready · no full dataset rewrites
planned
Get Updates

Built in SF.
Built for builders.

Early templates, release notes, and MatchArm-specific tooling drops. No spam — just the signal.

⭐ Star on GitHub 🔗 Live Demo