๐ฃ๐ฅ๐๐ฆ๐ ๐๐ซ: ๐ช๐๐๐ก ๐ฆ๐๐ ๐จ๐๐๐ง๐๐ข๐ก ๐ ๐๐๐ง๐ฆ ๐ฅ๐๐๐๐๐ง๐ฌ
A robot can perform well in simulation and still struggle with a simple task in the real world.
That's something I've been exploring while following
@PrismaXai .
At IROS 2026 in Pittsburgh, PrismaX demonstrated its work on robotic manipulation, using a task as simple as picking up fruit and placing it into containers.
But the interesting part is what happened before the demonstration.
๐ข๐ญ / ๐ข๐๐ฒ๐ฟ ๐ญ๐ฌ,๐ฌ๐ฌ๐ฌ ๐ฆ๐ถ๐บ๐๐น๐ฎ๐๐ฒ๐ฑ ๐ง๐ฒ๐๐๐
PrismaX ran thousands of simulated tests to understand where its robotic system might fail.
The goal wasn't just to make the robot complete a task once. It was to identify weaknesses before testing on real hardware.
๐ข๐ฎ / ๐ง๐ฟ๐ฎ๐ถ๐ป๐ถ๐ป๐ด ๐ ๐ฎ๐๐๐ฒ๐ฟ๐
One comparison from their reported experiments stood out.
โ 51% pickup success when training only the action component.
โ 89% when training the full model, including the vision and language backbone.
That suggests the robot benefits from adapting more than just its movement predictions.
The tradeoff? Training the full model takes more time.
๐ข๐ฏ / ๐ฅ๐ฒ๐ฎ๐น ๐ช๐ผ๐ฟ๐น๐ฑ ๐ฃ๐ฒ๐ฟ๐ณ๐ผ๐ฟ๐บ๐ฎ๐ป๐ฐ๐ฒ
Even with promising simulation results, the real robot didn't perform as well as expected.
The team also investigated inference latency and reported optimizations that reduced it substantially.
It's a useful reminder that robotics isn't only about having a capable AI model.
Hardware, timing, data quality, and the physical environment all matter.
๐ ๐ ๐ง๐ฎ๐ธ๐ฒ
What interests me about PrismaX is its focus on the practical challenges of physical AI.
Getting a robot to understand an instruction is one thing.
Getting it to execute that instruction reliably, on real hardware, under real-world conditions, is a completely different challenge.
And I think that's where some of the most important work in robotics is happening.
This demonstration is one specific task, with results reported by PrismaX rather than independent benchmarks. But the lessons about simulation, training, and real world execution are worth understanding.
The future of physical AI won't be decided by impressive demos alone.
It'll depend on how consistently robots can perform when conditions aren't perfect.