Exact-duplicate search misses mirrored, cropped, or filtered copies of your images — different bytes, same picture - voxel51.com/blog/image-copy-… That’s train/test leakage, wasted labeling spend, and licensing headaches waiting to happen. Voxel51’s fix: combine two signals that fail in different places. 🔹 pHash — reads pixel layout, breaks on mirrors/crops 🔹 CLIP — reads meaning, fooled by look-alikes 🔹 A rule you control to combine them (“any” vs “all”) 🔹 Every link shown as a graph, with evidence 🔹 Scored against ground truth with native evaluation Check out @DataScienceHarp's post! #ComputerVision #MachineLearning #OpenSource #AI #DataCuration

Failure mining in robotics: find one failure, surface every similar one - voxel51.com/blog/failure-min… Most robotics teams find failures by scrubbing hours of recordings by hand. This Voxel51 post shows a faster workflow in FiftyOne: * Start with one confirmed failure (a robot drops a cloth mid-fold) and mark it with a temporal tag * Use segment-level embedding similarity search to surface every visually similar moment across the dataset — not whole episodes, but 5-second windows * Embeddings find similar situations, not failures — a dropped cloth and a deliberately placed one look nearly identical. Scoping the search to episodes already flagged as failed cut candidates from 16 down to 2 * Review and confirm the real failures, then tag + save them as a reusable set for retraining or evaluation The result: a repeatable loop — find, search, narrow, confirm, tag — that turns a messy pile of recordings into a growing, labeled failure dataset. #Robotics #PhysicalAI #MachineLearning #ComputerVision #DataCuration #AI #FiftyOne #Voxel51 #RoboticsAI #DataFlywheel

Tokyo's most accident-prone intersection, turned into a dataset you can actually explore. Check out this lightning talk from @DataScienceHarp on the Hard Intersections Multimodal Dataset from Dynamic Map Platform Co., Ltd. - huggingface.co/datasets/dyna… * 6 synchronized camera views (full 360°) * 29 classes of road infrastructure — labeled entirely by models, zero manual annotation * Trajectory data at 200Hz * HD maps in 5 industry formats * ~35M-point aggregated LiDAR All in one interactive viewer with FiftyOne. No exporting. No jumping between tools. #ComputerVision #AutonomousVehicles #FiftyOne #OpenSource

VLA models don't learn from folders of images. They learn from episodes—vision, state, and action locked to one clock, frame by frame - voxel51.com/blog/lerobot-v3-… We loaded 11 LeRobot v3.0 datasets into FiftyOne so you can inspect them before you download. * Single Franka arms to full humanoids * Real teleoperation + pure simulation * Fingertip tactile sensing (20 streams) * From 10 to 27,500 episodes What we found digging in: * InternData-A1 is 100% synthetic—and looks photorealistic * droid_3d's depth streams silently drop (importer can't decode the dtype) * Licenses vary within a single export Look at your data before you train on it. Check out @DataScienceHarp 's post. #PhysicalAI #Robotics #VLA #LeRobot

When a racing drone's onboard navigation drifts, does it recover? Usually not - voxel51.com/blog/visual-iner… Explore the tii-ratm-drone-racing dataset in FiftyOne. 6 indoor quadrotor flights with onboard VIO recorded next to motion-capture ground truth. We checked whether the drift ever corrects itself: * 4 of 6 flights: worst error was the very last pose (no recovery) * RMSE alone gets it backwards, the flight that recovered had higher RMSE than ones that didn't * The fisheye tracker dropped ~45% of every frame across all flights * Worst flight drifted 1.66m RMSE The lesson: don't rank a SLAM stack on RMSE. Compute final-over-max, does it end on its worst moment, or correct itself? Why it matters → warehouse robots, inspection drones, any GPS-denied vision-only navigation. Check out @DataScienceHarp 's post! #PhysicalAI #Robotics #ComputerVision #SLAM #DroneRacing #VIO #FiftyOne #Voxel51

Join Voxel51 and Nebius at hubs.ly/Q04y0cx50 on Sept 26 for an afternoon of Physical AI talks and a hands-on robotics workshop - hubs.ly/Q04y0czK0 * A hands-on workshop taking real robot logs all the way to a trained detector * We’ll curate, embed, auto-label, and fine-tune YOLO11n in one sitting using Nebius Serverless AI Jobs * Explore multimodal robot data — camera, proprioception, and language on one synced timeline in FiftyOne. No ROS install required. Bring a laptop. Space is limited. Pre-registration is mandatory. #PhysicalAI #Robotics #ComputerVision #MachineLearning #FiftyOne #berlin

"Humanoid dataset" is doing a LOT of heavy lifting - voxel51.com/blog/humanoid-le… We pulled 5 humanoid + bimanual datasets into LeRobot v3.0 + FiftyOne. Same broad task — grab a thing, move it. But how does each robot describe that task? * State vectors range from 23 to 216 dimensions — a 9x spread * Same task category, totally incompatible representations * GR1T2 vs Reachy2: identical benchmark, identical cameras — state still differs by 8 dims, purely from kinematics * No coarse alignment applied — raw, as each team built them That gap is embodiment heterogeneity, and it's why a policy trained on one robot can't natively touch another's data. Each dataset ships with a live FiftyOne Space! Check out @DataScienceHarp 's post. #Robotics #PhysicalAI #VLA #MachineLearning #LeRobot #EmbodiedAI

The largest robot dataset in our new field guide: 637,498 trajectories. All synthetic - voxel51.com/blog/lerobot-dat… The two real, human-teleoperated ones? 4,258 and 1,804 episodes. That gap explains everything about VLA training right now. Real robot data doesn't scale like web data. We pulled 5 LeRobot v3.0 datasets into FiftyOne and compared how teams close it: * 2 real teleoperation sets (Franka + Flexiv arms) * MimicGen — re-stitches 1,250 human demos into new scenes * RoboTwin — LLM writes tasks, VLM catches failures, zero human demos * InternData-A1 — fully scripted, no demos at all The catch: only InternData-A1 reports actual zero-shot sim-to-real transfer — and it hasn't been independently reproduced. Synthetic data is scaling fast. Proof it transfers to real robots is still catching up. Check out @DataScienceHarp 's post! #Robotics #PhysicalAI #VLA #SyntheticData #MachineLearning #LeRobot #EmbodiedAI

Importing LeRobot datasets into FiftyOne with an agent! One sample = one episode. No parser required - voxel51.com/blog/lerobot-imp… The new fiftyone-lerobot-import agent skill: * Detects whether a dataset is still on the old v2.x layout * Auto-converts v2.x → v3.0 before import * Stops to ask before installing packages or creating a dataset * Runs as a thin layer over the open-source FiftyOne SDK — same import from a chat prompt or a checked-in script Chck out @DataScienceHarp's post! #PhysicalAI #Robotics #FiftyOne

Meta's Aria Digital Twin ships one episode as 7 files, 12.4 GB — and none of it opens in a viewer - voxel51.com/blog/aria-digita… Here’s hot to turn every ground-truth stream into one synchronized MCAP file for FiftyOne's multimodal viewer. ~300 lines of Python: * 12 channels — RGB, depth, segmentation, 3D boxes, skeletons, eye gaze, point cloud, trajectory, IMU — 181,444 messages, one file * Y-up vs Z-up: fix the axes or nothing lines up * 304 of 354 objects never move — log them once, don't redraw every frame * A clean run ≠ a correct run: add a physical-plausibility check message counts will never catch Check out @DataScienceHarp 's post! #PhysicalAI #Robotics #MCAP #FiftyOne

Your MCAP file writes clean. Validator passes. Message count matches. You open it in a viewer and nothing plays - voxel51.com/blog/mcap-author… That's hand-authored MCAP: the worst bugs stay silent until playback. * foxglove.CompressedImage written as JSON → every topic "Encoding unsupported" * Full epoch value in nsec → OverflowError * Calibration casing (D/K/R/P vs d/k/r/p) → 3 failed attempts on one message * Hand-rolled PLY parsers silently dropping faces → just use trimesh * Flipped sign or wrong calibration matrix → decodes fine, plays fine, still wrong * Two sensors, one schema, two clocks → half your channels land in 1970 The most expensive ones never throw an error at all. Check out @DataScienceHarp 's guide! #PhysicalAI #Robotics #MCAP #FiftyOne

Join Voxel51 and Nebius at hubs.ly/Q04w_0ZB0 on Sept 26 for an afternoon of Physical AI talks and a hands-on robotics workshop - hubs.ly/Q04w_0_P0 * A hands-on workshop taking real robot logs all the way to a trained detector * We’ll curate, embed, auto-label, and fine-tune YOLO11n in one sitting using Nebius Serverless AI Jobs * Explore multimodal robot data — camera, proprioception, and language on one synced timeline in FiftyOne. No ROS install required. Bring a laptop. Space is limited. Pre-registration is mandatory. #PhysicalAI #Robotics #ComputerVision #MachineLearning #FiftyOne #berlin

Robot data doesn't fit in a folder of images - voxel51.com/blog/mcap-datase… A driving rig fires its camera, LiDAR, IMU, and GPS on four different clocks. Force that into "one image = one label" and you lose the timing or explode your sample count. So we repackaged 10 robotics datasets as MCAP episodes you can explore natively in FiftyOne: * 89 episodes, 6 domains — driving, SLAM, aquatic, off-road forest, social navigation, and one synthetic set. * One sample = one full recording — cameras, LiDAR, radar, sonar, IMU, GPS decoded live and scrubbable in sync. * Honest about labels: only 3 of 10 ship real object detections. We flag exactly which — plus every license quirk. Check out @DataScienceHarp 's post! #PhysicalAI #Robotics #MachineLearning

The smoothest robot episode in the batch also failed the task - voxel51.com/blog/motion-smoo… We scored all 127 episodes of a cube-in-bowl task from RoboLab-EgoX on motion smoothness. The result: * One episode ranked 8th-smoothest of 127 — and still failed. Zero flags, no anomalies, every check clean. * Metrics like SPARC and LDLJ measure how the arm moved, not whether the cube landed in the bowl. A fluid path to the wrong place scores as flawless. * Lesson: score for triage, not gates. Smoothness tells you which episodes to watch — never which ones passed. Check out the post by @DataScienceHarp #PhysicalAI #Robotics #MachineLearning