PrismaX

Where humans, robots, and data come together to advance intelligence. 🦾 Backed by @a16z. Host of @RoboCon_AI 🤖
🔗

[PrismaX @PrismaXai]

With 24 GB laptop GPUs getting harder to find, we squeezed our π0.5 rollout stack from about 22 GB of VRAM to 11.2 GB.

New from @castorhat: the two fixes, and the 16 GB cards it runs on now. https://t.co/N1oynHK4MH

[PrismaX @PrismaXai]

Watch the full panel with @vivekgr92 of @agilityrobotics, @sharonal_lee of @moonlake, and @castorhat of @PrismaXai:
https://t.co/fTtuyiR7TQ

[PrismaX @PrismaXai]

"Imagine if you're testing a robot policy on a hundred thousand dollar humanoid robot and it breaks. Then that's it."

@vivekgr92 of @agilityrobotics earlier this month on our RoboCon Physical AI Salon panel, From Lab to Deployment: https://t.co/fTtuyiR7TQ

His point:

[PrismaX @PrismaXai]

We test this way at PrismaX too. As @castorhat put it on the panel, failures in sim are trivial to recover from. Failures on hardware cost you weeks.

[PrismaX @PrismaXai]

Both repos are open source:
yamulator (sim) https://t.co/Gpf69r6Npi
prism-pi (rollout stack) https://t.co/Syfh1SiQfJ

We'll have another demo running at #CoRL2026 in Austin. Come see it at Booth #1100 👀

[PrismaX @PrismaXai]

This is our π0.5 demo running live at #IROS2026 last month on a different YAM rig than the one its training data came from.

It had to work in a crowded and loud convention hall none of us had seen before, with people walking around the arms all day.

@castorhat broke down what it took to get it working in our new three-part series 👇

[PrismaX @PrismaXai]

Part 3 is about memory. Two fixes took peak VRAM from about 22 GB to 11.2 GB so it now runs on 16 GB cards like the RTX 5080.

https://t.co/APn5ZuxRh0

[PrismaX @PrismaXai]

@artzarn Thanks for the breakdown!

[PrismaX @PrismaXai]

@TheHumanoidHub The API abstraction changes who can deploy a policy. Natural-language prompting without task-specific data significantly lowers setup cost.

[PrismaX @PrismaXai]

@lukas_m_ziegler Super impressive.

[PrismaX @PrismaXai]

@AGIBOTofficial Moving the benchmark in that direction is important. Let's see what deployment actually means in practice.

[PrismaX @PrismaXai]

@ScaleAILabs @huggingface Big news!

[PrismaX @PrismaXai]

Thank you to Boston Venture Connect for hosting, and to @Microsoft, Midas Capital, and COGITO for partnering.

The work starts after the demo. Good to spend an afternoon in Cambridge with people building past it.

[PrismaX @PrismaXai]

A demo isn't the milestone. A customer who's happy because you solved a real problem is.

That was @castorhat's first point on the "Beyond the Demo: Building AI Businesses" panel at Boston Venture Connect's Greater China AI Showcase in Cambridge last week.

[PrismaX @PrismaXai]

Check existing hardware before you build your own.

Custom hardware still costs serious time and money, and off-the-shelf robots and platforms often do the job. Our IROS demo ran on two off-the-shelf YAM arms.