Robots Digest 🤖

Follow @RobotsDigest for latest in Robotics, Humanoids, and Hardware + AI.

[Robots Digest 🤖 @robotsdigest]

LeRobot docs: https://t.co/mq04lfj4fa
Blog: https://t.co/yu3wpoiilS

[Robots Digest 🤖 @robotsdigest]

Real world data is scarce and expensive to collect in the volume a generalist robotic policy needs.
Runway just destroyed this bottleneck with Praxis 1.

This open weight world action model turns massive video pretraining into physical control, giving robots an enormous head start on physical plausibility without relying purely on action data.

[Robots Digest 🤖 @robotsdigest]

Praxis-1 is the strongest world action model available: #1 on NVIDIA’s RoboLab-120 leaderboard, and ahead of every open-weight VLA on the board, including π0.5 and GR00T N1.6.

[Robots Digest 🤖 @robotsdigest]

Project page : https://t.co/zLT0oZzFOQ
Github : https://t.co/lpKSe8mxiP
Paper : https://t.co/h4r21dg4qY

[Robots Digest 🤖 @robotsdigest]

What if 3D was just... an image?

UniFlow (CoRL 2026, UW + Toyota Research Institute) feeds a robot policy a pointmap: for every pixel, the (x, y, z) point that pixel sees.

Because it’s the same shape as the RGB frame, an off-the-shelf pretrained 2D encoder can read geometry too.
Geometry inherits 2D pretraining.

[Robots Digest 🤖 @robotsdigest]

A small policy with no robot data pretraining just beat the large scale pretrained VLA π0.5.
UniFlow learned 4 real bimanual tasks on the YAM platform (Plate Handoff, Plate Placement, Cup Stacking, Spoon Sort) from only 191 demos, about an hour of teleoperation.
The biggest gap is on the thin plate handoff. π0.5 often misses the edge and drops it, while UniFlow passes it between arms and racks it. UniFlow also had the best average on 5 RoboTwin bimanual tasks.
Its edge shows up on tasks needing precise 3D placement.

[Robots Digest 🤖 @robotsdigest]

Paper: https://t.co/2wftc9CZfE
Project: https://t.co/MYvmBAkwCq
Code: https://t.co/XterbmbJ9L

Led by Billy Yan (@BillyYYan, applying to PhDs this cycle) with @Isaac1385 @KellyLee1742558 @1bcx3 @Kmang0s @patel_kanad @DhawalKabra @neeerpatel @kevin_y_wu @LerrelPinto and Irmak Guzey.

[Robots Digest 🤖 @robotsdigest]

Five human demos. Under five minutes. A robot learns a brand-new task, from recording to deployment.

That's Steertuning from NYU: 71% average success across 12 new tasks, including kitchens in new homes.

[Robots Digest 🤖 @robotsdigest]

Chained with navigation, one sentence can run a full task.
Ablations favor our architecture and object-based demo selection, and success grows with compute.
Code is open, and weights and dataset are coming soon.

[Robots Digest 🤖 @robotsdigest]

Author : @huksyn
Paper: https://t.co/99VuQs1W7y
Code: https://t.co/pqwTlc850u
Project Page: https://t.co/zs3VCgEHuR

[Robots Digest 🤖 @robotsdigest]

Apple researchers just introduced
HALO - Perceptive Humanoid Motion from Simulation
Trained entirely on synthetic data from 1,000 simulated homes.

It achieved 91% success across 360 physical trials, with zero-shot deployment in 6 unseen real-world environments , all fully onboard.

[Robots Digest 🤖 @robotsdigest]

The scale is impressive:
1,000 simulated homes
~300 hours of synthetic data
91% success across 360 real-world trials
The robot learned in simulation. The test happened in reality.

[Robots Digest 🤖 @robotsdigest]

paper:https://t.co/M3woTh8lqr
website:https://t.co/e9mgsIC0dv
code:https://t.co/bKCZqOePqx

[Robots Digest 🤖 @robotsdigest]

A robot's palm feels the load. Its fingertips feel the deformation.

RevoSim brings both into one simulator: whole-hand pressure sensing plus optical touch at all five fingertips.

What always breaks is making these two kinds of touch work together

[Robots Digest 🤖 @robotsdigest]

The demos: rotating a ball using fingertip depth, and grasping with palm pressure plus fingertip depth.

Two complementary signals: local contact shape and contact beyond the fingertips