We're all-in on a new product. @HapticLabsAI
is now AI for construction and development, a land intelligence engine with an agent on every parcel.
Land diligence that used to take 3 months of legal, zoning, permits, earthworks, economics we now compress into 2 days.
Paying customer #1 is live with us. If diligence is still the bottleneck, come talk to us.
At @hotchipsorg for the next 2 days. Shoot me a DM if you’re here! Excited to learn more about what’s coming out in the chip world and especially how EDA is evolving to meet the new demands
What problems and bottlenecks keep coming up in physical AI research?
We analyzed 64 episode transcripts of the brilliant @RoboPapers podcast to see which bottlenecks robotics researchers mention the most.
https://t.co/NTaG4WqlOf
In March, we mapped the pain across 64 RoboPapers episodes ... this time, we mapped the surprises that most changed our thinking. Here are eight of the surprises after scanning those same episodes.
TL;DR: better systems around a model often beat better models.
https://t.co/ou1yIzZxzh
MolmoAct2 (so101 ft) does well in some basic pick and place tasks and runs smoothly on @HapticLabsAI's inf. infra. A bit surprised that it's not been able to distinguish colors or shapes (e.g. doesn't distinguish between a yellow or a blue cube, or between a cube and a cylinder)
Robot training pipelines do not always fail loudly
We harvested 267 public datasets tagged 16-DOF OpenArm. Only 45 merged cleanly: 652 episodes and 583k frames
Read about how we mapped what broke and had an agent curate the corpus and fine-tune Pi0.5 ↓
https://t.co/ePKAhxgV05
Three open robot models dropped in three days. @LeRobotHF 0.6.0 on Tuesday, with six benchmarks to evaluate a policy before it touches hardware. LingBot-VLA 2.0 the same day: a 6B open checkpoint trained on 60,000 hours of real robot data. @Nvidia GR00T 1.7 this morning, on a 2B backbone.
Where @HapticLabsAI steps in: we do everything after the download. Shrinking a policy to one robot's compute budget, serving it at control-loop rates, and managing a fleet where some inference runs on the robot and some in the cloud.
As we build out these AI Factories at GW scale; procuring space, power and GPUs is just the first step. Maximizing the Return on Investment (ROI) requires continuous optimized software and hardware availability and ensuring we deploy and bring-up the AI Factories in days and not after months of debugging.
Robot policy deployment should feel more like shipping software.
Today it doesn't. A policy that works in a demo still has to run on the robot, hit latency targets, survive real networks, and be managed across a fleet.
That's what we've built at Haptic: the deployment and inference layer for robot policies.
This week @pham_blnh got a robot policy running inside a browser tab. It works because that policy (ACT) is about 80M parameters, small enough to live in a tab.
It's a great on-ramp. It also quietly shows the deployment problem we work on every day. The policies that run in robots are larger: generalist VLAs are around 50 times bigger and still have to hold 20Hz on an 8GB computer bolted to the arm.
The 10TB to 2TB re-encode is genuinely nice engineering. Credit to LeRobot for the unglamorous part. A lot of the companies we have worked with have cited this as a big pain.
I couldn’t believe when I saw this at SFO. The obvious take is this is some crazy weak energy.
My half-baked question: is this a sign that tech is forgetting how to market to human consumers? Did that muscle atrophy as B2B dominates, AI slop
permeates, people sell to agents, and old ideas like brand give way to algorithms.
Seems like people who know the old school can dominate in this new world.
We are super excited to share with you our initial release of Lucky Engine. We are building a robotics engine from the ground up to be what we wished we could find in a simulator before
China's Unitree Will Dominate Global Robotics
The Fastest Iteration Cycle In Next-Gen Robotics Should See Unprecedented Acceleration
https://t.co/0HKSayvvTK