Blog

Technical guides on robotics data pipelines, dataset formats, and training data infrastructure.

July 23, 2026 9 min read

Your Robot Doesn't Know Where the Table Is. Neither Does Its Training Data — We Checked.

We audited seven representative LeRobot datasets on Hugging Face — 9,694 episodes combined. None declared camera calibration, six declared no depth stream, and none carried a versioned spatial artifact. The 3D signals the newest VLA papers rely on are missing from the data.

Data Quality Spatial Context LeRobot Calibration
June 5, 2026 6 min read

Is Your Synthetic Robot Data Actually Diverse? We Put a Number on NVIDIA's.

We ran our embedding diversity check on NVIDIA's IsaacSim synthetic manipulation data: near-zero visual redundancy — it passes the test scripted-sim datasets fail. Plus a false positive we caught and fixed in our own QA.

Synthetic Data Data Diversity NVIDIA QA
May 31, 2026 7 min read

We Put a Vector Layer Over 5 Popular Robotics Datasets. Here's What That Unlocks.

A DINOv2 embedding over every episode, run through our dataset CI. Four scripted-sim datasets collapse to a single visual cluster; libero splits into ten. The diversity and tail-episode signals never show up in rule-based QA.

Embeddings Data Quality DINOv2 Curation
April 23, 2026 8 min read

We Audited 5 Popular LeRobot Datasets. 4 Ship Stats That Produce Inf.

We ran traceplane check on five of the most-downloaded LeRobot datasets on HuggingFace. Every single-task dataset ships normalization stats that produce Inf the moment your dataloader uses them — silently.

Data Quality LeRobot Normalization Datasets
April 8, 2026 10 min read

GEN-1 Proved Human Data Trains Robots. Here's the Infrastructure You Need.

Generalist AI achieved 99% success rates using 500K hours of human activity data and zero robot data. Here's the data infrastructure required to replicate this approach.

Human Data VLA Data Pipeline Infrastructure
April 7, 2026 12 min read

We Audited 10 Popular Open-Source Robot Datasets. Here's What We Found.

We ran automated quality checks on 10 widely-used robotics datasets including Bridge V2, Open X-Embodiment, ALOHA, and LeRobot datasets. Every single one had issues that could silently degrade your policy.

Data Quality Datasets Imitation Learning LeRobot
April 4, 2026 10 min read

Automated QA for Robot Trajectory Data: A Three-Layer Framework

Why "record more demos" doesn't fix training failures. A practical framework for structural, kinematic, and semantic quality checks on every episode — automatically.

Data Quality QA Imitation Learning
April 4, 2026 8 min read

How to Convert rosbag2 Data to LeRobot Format

A practical guide to converting ROS 2 bag files (.mcap, .db3) to HuggingFace LeRobot's Parquet + MP4 format for policy training. Covers timestamp alignment, video encoding, and schema mapping.

rosbag2 LeRobot Data Conversion