DataSnack

Crumb, by DataSnack, is an independent third-party black box for robots.

[DataSnack @data_snack_ai]

Physical-AI teams: which real operating environment is hardest to secure for field evaluation—warehouse, retail, restaurant, office, industrial, or something stranger? DataSnack helps source and coordinate supervised testing locations around the requirement.

[DataSnack @data_snack_ai]

A food delivery robot at ASU reversed into a 73-year-old woman. Twice. The only technical record of what happened is the robot company's own telemetry — produced by the party with the most to lose. That's not evidence. That's a press release. #robotics

[DataSnack @data_snack_ai]

EU just expanded vehicle event recording mandates. When a car is in an incident, there's an independent record. When a robot clips a worker or wrecks a pallet, the only record is the robot company's own telemetry. Same liability problem. Cars solved it. Robots haven't.

[DataSnack @data_snack_ai]

A robot vacuum put a 25-year-old in critical care with burns on 75% of his body. Active investigation. Forensic testing underway. Who controls the data on what that device did? The manufacturer. That's not evidence. That's a company investigating itself. #physicalai

[DataSnack @data_snack_ai]

Berkshire, Chubb, and Travelers are seeking to exclude AI liability from their general policies. Commercial robots still have zero independent incident recording. No neutral evidence when things go wrong — just the robot company's logs saying everything was nominal.

[DataSnack @data_snack_ai]

Berkshire, Chubb, and Travelers are seeking approval to exclude AI-related damage from general liability policies. When a robot causes an incident and the only record is the robot company's own logs — that's a "decision nobody can explain." #robotics

[DataSnack @data_snack_ai]

Berkshire, Chubb, and Travelers all filed to exclude AI from commercial liability policies. They can't independently verify what happened when an AI system causes damage �� only the self-reported data from the party with the most to lose. That's the gap. #robotics

[DataSnack @data_snack_ai]

Apptronik built a Robot Park to train humanoids for real-world environments. Good instinct. Bad substitute. The variables that break pilots - staff schedules, floor changes, wifi dead zones - don't exist in any park. They exist in the facility. #physicalai

[DataSnack @data_snack_ai]

Evaluating your manipulation policy on held-out synthetic data doesn't measure sim-to-real transfer. It measures whether your test split differed from your train split. Those are not the same question.

[DataSnack @data_snack_ai]

Per-frame QA at capture: depth confidence, ARKit tracking state, hand detection confidence. Untracked frames get identity quaternion + confidence=0 — flagged, not dropped. Filter in your dataloader. pose_consistency.json in every export. Know which frames to trust before you...

[DataSnack @data_snack_ai]

If your evaluation dataset doesn't ship with per-frame QA metadata, you're guessing about data quality at test time. pose_consistency.json, depth confidence, hand tracking confidence — the signal that tells you which frames to actually trust.

[DataSnack @data_snack_ai]

Your manipulation policy's eval set probably came from the same simulator that generated its training data. That's not testing sim-to-real transfer, it's testing whether you remembered to hold out a few episodes.

[DataSnack @data_snack_ai]

Every household session gets a pose_consistency.json before it ships. Static mesh gets re-rendered at every claimed camera pose, diffed against measured depth. Drift past threshold flags the frame instead of letting it ride into the eval set quietly.

[DataSnack @data_snack_ai]

Evaluation data quality isn't binary. datasnack/brian_does_cleaning ships per-frame depth confidence, hand tracking confidence scores, and pose_consistency.json with every session. Know which frames to trust before you run your policy through them.

[DataSnack @data_snack_ai]

NVIDIA's world-action models post leans on DROID: 50k demos on Franka arms, in labs. Pretraining doesn't dissolve the distribution gap — it defers it to fine-tuning, which still needs demos from the warehouse your robot deploys in. Your policy knows the lab.

[DataSnack @data_snack_ai]

iPhone + ARKit LiDAR: RGB-D, camera poses, IMU, scene mesh, LeRobot v3. Real warehouse manipulation data in days, sub-$100 hardware. Most teams find out iPhone depth is good enough from the pilot data, not the spec sheet.

[DataSnack @data_snack_ai]

Catching a depth dropout while the operator's still at the station costs you a 30-second re-capture. Catching it three weeks into training costs you the whole batch. Real-time QA on capture day isn't a nice-to-have, it's the difference.

[DataSnack @data_snack_ai]

RoboNet: controlled lab, North American facility, staged setup. Your robot is deploying in a Lagos port warehouse. Zero public datasets from that environment. That's not a fine-tuning problem.

[DataSnack @data_snack_ai]

95% on sim rollouts. In a real warehouse, that poly bag deforms in ways your physics engine never modeled. The policy knows the sim. It doesn't know the material.

[DataSnack @data_snack_ai]

Physical AI is deploying across Africa and LatAm this year. The training data for those models was collected in the US and EU. That gap doesn't close with more epochs — it closes with data that doesn't exist yet.