Vector Wang

XLeRobot 0.3.0 Showcases Open fridge, get drinks, fill ice, wipe table, clean room, take care plants and cats... All for 660$, fully open-sourced, based on HF LeRobot. Teleop with Joy-con, or RL/VLA. Assembly kit ready for purchase soon Stay tuned! github.com/Vector-Wangel/…

I still hold my opinion that action-conditioned world model fits the core definition of "model": s_t+1=f(s_t, a_t). However, there's no model that can be accurate, even in real life, all we have are blurry "dreams"

XLeRobot can finally self-start and self-calibrate, thanks to the contributions from @QILIU9203 @IsaacSin12. This way even we are miles away from home, we can still turn XLeRobot on directly.

@Ryan_Resolution @IsaacSin12 @LeRobotHF @huggingface So if connected to a Jetson and the Jetson is remotely controlled with SSH, the robot can self start without any manual work?

@IsaacSin12 @LeRobotHF @huggingface @VectorWang2 what do you think? 😉

Auto-calibration @LeRobotHF update v1 had a ~30% failure rate from Feetech servos stalling and restarting. Also got feedback it was moving too fast and are damaging servos. Fixed both. Now works reliably on leader + follower arms and mounted on the XLeRobot

@jinwanxin Really cool work! Can it be directly included into newton?

This is a project we’ve been working on for a long time, and today we’re incredibly excited to share ComFree-Sim — a GPU-parallelized ANALYTICAL contact physics engine built to be lightweight for faster, scalable, dense-contact simulation. It’s a drop-in alternative to MuJoCo (same API) that skips iterative solves to keep dense-contact compute flat. 👇 ⚡ >2× higher throughput in dense-contact scenes vs. MuJoCo Warp 🎯 Comparable, highly tunable contact dynamics fidelity 🤖 Low-latency sim-predictive control for dexterous manipulation, humanoids, and more Faster, learnable physics unlocks more for closed-loop physical intelligence. Website, paper, and videos: irislab.tech/comfree-sim/ 🚀 ⏳ ComFree Warp drops this weekend. ComFree Jax coming soon! #Robotics #EmbodiedAI #Simulation #PhysicsEngine #DexterousManipulation #Humanoids #MuJoCo #GPU

@kevin_zakka Are you considering adding newton also as the backend along with mjwarp? It could greatly expand the application directly.

Coming soon to mjlab: heterogeneous worlds, aka every world gets its own object 👀

@ChongZitaZhang Got it, then really look forward to adding this work also into newton

@VectorWang2 Newton can work with arbitrary solver that is written with warp

This is why we need Newton

Exactly the pipeline I imagine with Foundation Stereo and SAM2, now with FFS, TiPToP can directly become realtime deployable!

@JieWang_ZJUI Can’t wait to try it on my new XLeRobot. With fast foundation stereo this could run in real time and close the loop.

TAMP vs End2End, which one is better? Check out our latest research ablating these two on tabletop pick-and-place setup, it turns out the SOTA foundation models provide very good prior that solve this task family. Please enjoy the download-and-play TipTop from MIT folks!

@AntonioSitongLi why not use this as the robot head lol. The current one is too big

I made a little guy

@rovers_x @bowenwen_me github.com/Vector-Wangel/…

@VectorWang2 @bowenwen_me Cool! Is there a link for me to try?

Fast Foundation Stereo + SAM2 basically = zero shot Foundation Pose😂 No CAD model, no object image, just click the target. Run directly on my 3070 at 13fps, with a 30$ stereo camera (calibrated in 10min) Thanks @bowenwen_me for his contribution to the community!

@macjshiggins @bowenwen_me github.com/Vector-Wangel/…

@VectorWang2 @bowenwen_me This is awesome. Do you have a GitHub link?

Here’s the code (just randomly name it FFS-Pose): github.com/Vector-Wangel/…

@iamRezaSayar @bowenwen_me SAM3 is bigger and actually runs much slower than SAM2. SAM2 is already doing a great job at real time.

@VectorWang2 @bowenwen_me nice! i wonder if SAM3 would lead to any improvements here 👀

@macjshiggins @bowenwen_me Yeah basically, use SAM 2 to track segmentation from RGB, map it to the point cloud given by FFS, then just a simple bbox algorithm.

@VectorWang2 @bowenwen_me Is it literally just those two libs wired together? That’s convenient

@real_vaishak @bowenwen_me 138RMB（20$）if bought in China, in the US it’s around 60$

@VectorWang2 @bowenwen_me Wow! what's the $30 cam?

This is a really great function. Just like human stretching their muscles, robot also needs to “stretch” itself. Hoping to see the robot one-click start soon!