This Spot rolls a 15 kg tire heavier than its arm can lift. π
RAI Institute took it out of the lab.
β Sees the tire with its own depth cameras, no mocap
β Knock it over mid-roll, it stands it up
β Skatepark, basketball court, garage
Credit: RAI Institute (@rai_inst)
Show this robot a task once. It learns it instantly. π€
Machenta just came out of stealth with a $40M seed.
β One demo, no further training
β Trained entirely in simulation, zero real-world data
β Same model runs on six different robots
Credit: Edward Johns (@Ed__Johns)
Google's new AI builds worlds you can steer live. π
It's Genie 4, from Google DeepMind.
β Prompt mid-scene to change the weather or add people
β Makes its own sound: ambient, music, dialogue
β Testers now, Google AI Ultra soon
Credit: Google DeepMind (@GoogleDeepMind)
A drone chasing a runner through the woods with no GPS and no map π²
LiDAR-inertial odometry tracks its own position
Every scan feeds a live 3D occupancy grid
Runner ducks behind a tree? It predicts where they went
All of it computed onboard
Credit: Aniket Bera (@anikbera)
OpenAI's new model built a 3D tram from one photo, running the 3D app itself. π
It's GPT-6.1 Terra.
β Matches GPT-6 Astra on agent tasks at 1/10 the price per token
β Over 3x faster than Astra and Sol
β Live now in ChatGPT and Codex, paid plans
Credit: OpenAI (@OpenAI)
Figure went from a humanoid on a safety tether to a wall of robots making beds. ποΈ
Brett Adcock just put all 3 generations of Figure's humanoids in one clip.
Watch the ending: a grid of robots doing chores in homes.
Credit: Brett Adcock (@adcock_brett)
We've all typed --dangerously-skip-permissions. π¬
Give your agent a box. It's called tryboxi.
β One microVM per agent, on your Mac
β Claude Code, Codex, Cursor
β Ready in 20 to 50 ms
β Secrets never enter the VM
β Free and open source
Credit: Karthik (@karthikponna19)
Block just built a voice AI that never read a single word. ποΈ
Dessa learned to talk only by listening to people.
β No text, no pretrained models
β Listens and talks at the same time
β Runs locally on a MacBook Pro M4 Max
Credit: Block (@blocks)
A humanoid riding on the end of an industrial robot arm. ποΈ
LimX Dynamics mounted TRON 2's upper body on a giant arm. The arm brings reach, TRON 2 does the hand work.
Firing a nail gun, building formwork, running two drills at once.
Credit: LimX Dynamics (@LimX_Dynamics)
This bee started as a yellow cube in Blender. π
Higgsfield blocked out the scene, then Seedance 2.5 rendered it in native 4K.
Top: the final video.
Bottom: the blockout it followed.
Credit: Higgsfield (@higgsfield_ai)
Film a robot on your phone. An AI agent builds the sim. π±
It's RoboSimStudio, from UC Berkeley and NVIDIA.
β Claude or Codex builds the twin in about an hour
β Models sponges, cables and cloth
β Train in sim, run on the real robot
Credit: Kaiyuan Eric Chen (@keplerccccc)
This robot runs its whole conversation on a 4GB model. π€
NVIDIA just released Nemotron 3.5 Bolt E4B, a small open model.
β Bolt handles chat and tool use
β NVIDIA Cosmos handles vision
β Both run locally on a Jetson Orin Nano 8GB
Credit: NVIDIA AI (@NVIDIAAI)
This humanoid climbs ladders and picks its own holds. π§
ETH Zurich taught it where to make contact with just 13.2 minutes of human motion.
β One policy for every skill
β Plans contacts from LiDAR and a goal
β Falls, gets up, keeps going
Credit: Junzhe He (@JayHe748646)
Your coding agent can now use your whole computer. π±οΈ
Cursor just shipped Computer Use.
β Clicks and types in any app, not just the browser
β Drives the iOS Simulator to test what it built
β Kicks in on its own when a task needs it
Credit: Cursor (@cursor_ai)
These robot hands close in about 80 milliseconds. π₯
Meet the tendon-driven hands on RAI Institute's AthenaZero.
β Heavy tools or fragile eggs
β 11 sensors per fingertip and palm
β Driven by motors in the torso
Teleoperated in this clip.
Credit: RAI Institute (@rai_inst)
This humanoid reads your hands. No voice commands. π€
XPeng's IRON shakes hands, bumps fists and high-fives in real time.
Switch to a high-five mid-handshake and it adjusts on the fly.
VLA 2.0 re-reasons on what it sees every 0.2 seconds.
Credit: XPENG (@XPENG_Global)
Ideogram just dropped its first vector model. βοΈ
Type a prompt, get a clean, editable SVG.
β Smooth gradients, sharp typography
β Turn any image into a vector
β Swap the palette, the gradients follow
β Live in Ideogram and the API
Credit: Ideogram (@ideogram_ai)
Baymax isn't just a kids' movie anymore. This humanoid feels hugs π
The skin is an inflatable Baymax costume.
Inside, 116 ToF sensors watch the fabric at 20 Hz.
An MLP predicts contact-free skin from joint angles. Dents = contact.
Credit: DGIST IRL (https://t.co/QwjyWvHtHi)
Jensen Huang got down on one knee for robot ducks. π¦
Meet Pollen Robotics' Microducks, on stage at his NVIDIA keynote.
25 cm tall, 15 motors, $399.
They walk, roller skate and get up after falls. Skills are trained in simulation.
Credit: Pollen Robotics (@pollenrobotics)
Unitree's H2-X jumps off a 4 meter platform and lands it π€
It also kicks and punches through junked cars. Unitree says it's real footage, no speed-up.
Later it lifts a car to free a trapped dummy.
Claimed standing endurance: about 8 hours.
Credit: Unitree (@UnitreeRobotics)