BTS from the lab. Calibrating our stereo cams by hand was landing around 0.7 px reprojection error. Good is under 0.5. So the team put a YAM arm on it. It moves the camera through a fixed set of poses in front of a ChArUco board, nobody touching it. Same poses every run, so the capture set is well-conditioned instead of depending on who held the camera that day. ~0.3 px now. Two minutes a rig. Nobody asks about calibration. Everything downstream inherits its error.

zero shot attempts.

working on calibration multicam system. hardware used is robocap from @BitRobotNetwork this plot shows all the six cameras mapped in one shared reference frame. here, LF (left-front) is taken as the master cam and all are transforms are against it. I'm seeing some constant rotation between ind. stereo cams i.e left and right of front pair (LF, RF) not sure if this is a hardware level issue, calibration issue or some damage happened e.g drop. I first tried doing the calibration using the AprilGrid tags std. 6x6 as recommended by kalibr but it failed massively on highly distorted fisheye cams of bitrobot. we observed cushion distortion effect on the rectified images. it's basically our k distortion coeff. wasn't solved properly. (might be that I didn't had full frame coverage of the aprilgrid in the camera frame, it's a WIP) charuco board has been used for these output plots.

been working on UMI's for the past week. how's the output?

I think we need a VR headset understanding to build sota egocentric data capture rig

Just us being hands on. Something’s coming. Watch this space.

work in progress! what's the preference for UMI data collection? parallel jaw based gripper or these?

what's the most efficient method of tracking in umi's without using VR trackers, mirrors?

atlas

interesting to see how most slam algos fail on egocentric videos collected in wild

What do you do with 16TB of pure robotic manipulation data? 🤖 You open source it on @huggingface Simple World Lab just dropped HiFi-UMI-2K - an absolutely massive dataset pushing the limits of high-fidelity manipulation. The entire 16TB is already available in the LeRobot format, meaning you can plug this massive scale straight into your pipelines with zero friction. Dataset: huggingface.co/datasets/simp… Paper: huggingface.co/papers/2607.2…

working on a smaller world action model (wam) which relies on classical algorithms and chunks of pre-exsisting models. here, i'm able to mask out the object of interest and track it during the whole manipulation duration. wam predicts future state both in pixel and action space. 14B params model outputs the future frames. i'm trying to cut the huge model part out. data source: @AGIBOTofficial paper followed: dreamzero and dream dojo from nvidia dreamzero0.github.io/ arxiv.org/abs/2602.06949

is it okay to compress videos from their original source using a neural encoding method and then use them for model training?

piped.video/watch?v=3Y8aq_of… great presentation! defines a clear research direction.

on ACT's human demos, 4× more joint-space variance sits in the null space of the end-effector jacobian than in task space. policies regressing joint angles are spending capacity on the demonstrator's elbow.

world models.