Topic hub
Physical AI & Robotics
Robot training data, teleoperation and egocentric collection, dataset formats, simulation, and policy training for machines that act in the world.
All posts · 13
PHYSICAL AI & ROBOTICS01
LeRobot v2.0 to v3.0 Dataset Conversion Takes Two ReleasesA v2.0 LeRobot corpus needs two releases to reach v3.0: 0.3.3 for the first rung, 0.6.1 for the second. The first rung cannot run offline as shipped.
PHYSICAL AI & ROBOTICS02
LeRobot Dataset Merge: What Breaks and How to CheckThe documented merge command raises before it opens a file, and a merged q01/q99 has never been a real quantile at any LeRobot version. Check yours in one pass.
PHYSICAL AI & ROBOTICS03
Removing Episodes From a LeRobot Dataset: Which to CutLeRobot v0.6.1 deletes episodes in one command. Deciding which to cut is the harder half: four proxy criteria, one real label, and a written bar.
PHYSICAL AI & ROBOTICS04
Isaac Lab vs MuJoCo vs Genesis for Robot Training DataOne MuJoCo point release moved a LIBERO task from 98% to 12% success on an unchanged checkpoint. How to choose a simulator and pin its physics.
PHYSICAL AI & ROBOTICS05
Open X-Embodiment Dataset License: No Single Answer1,876 of the 2,000 most-downloaded LeRobot datasets read apache-2.0. That tag is a Python default argument, not a licence anyone declared.
PHYSICAL AI & ROBOTICS06
Joint Space vs Task Space: Which Robot Action Space to Use13,000 real rollouts put the delta anchor ahead of the joint versus task choice, and DROID ships eight action columns in one episode.
PHYSICAL AI & ROBOTICS07
Where to Put the Cameras for Robot Imitation LearningShift the camera 45 degrees and ACT success drops from 0.83 to 0.13. Three groups measured the same collapse. Rig layout, resolution and the fixes.
PHYSICAL AI & ROBOTICS08
SmolVLA Fine-Tuning Returns 0% Success: What to Check FirstEight checks in cost order for a SmolVLA fine-tune that trained cleanly and rolls out at 0%. Start with lerobot-replay, not with a bigger dataset.
PHYSICAL AI & ROBOTICS09
How Many Trials Does It Take to Evaluate a Robot Policy?Seven successes in ten trials gives a 95% interval of 39.7% to 89.2%. That demo cannot rank two policies. Here are the trial counts that can.