Vibe Coding Robotics Part 9 First demo with Gemini 3.1 Pro Industrial robot arm IK simulation Design generated with Nano Banana

Compared to Gemini 3 Pro, I think 3.1 version has noticeably better than version 3 at understanding how to place items in a 3D space. For the other UI tasks I didn't notice a significant difference.

In case you missed it, here is the last part of the Vibe Coding Robotics series where I built a hexapod simulator with Claude Opus

I also have a series where I try to make games with AI. Here is the last part of that series. Will try to make something new with new Nano Banana version next.

So Is this like a web app or something? Are all of these numbers, controls and values actually tied to / mean something or are they mostly for aesthetics purposes for now? 😁

Yeah, it's a web app. The dials for the joint angles represent actual rotation values from the IK solution. The velocity metrics and force readouts are based on a noise function, so those aren't tied to anything

How much of this is just a video file with no accurate metrics?

How is this made? I would like to learn.

more details about my process are in this post

Looks pretty cool

how do you feel gemini 3.1 performs against opus for 3d?

For physics simulations like this, I feel Gemini is better than Opus. I tried building a terrain editor for games and CAD software using both Gemini and Opus. For those tasks, Opus performed better. So I would say that as the codebase becomes larger, Opus tends to perform better overall.

this is where llms get interesting - when they touch the physical world, the failure modes are different and the rewards are bigger. software-only ai wrappers are a crowded space. llms + robotics? still early, big upside.

AI is becoming a real robotics co-engineer.

Neat, what's the physics model?

🦾🫡

@GoogleDeepMind is all about jus simulation thanks to @demishassabis

kinda coool

now ai getting interesting

Inverse kinematics is where the spatial reasoning difference actually shows up. Most benchmarks test text generation, not whether the model can hold geometric constraints in mind.

Gemini 3.1 Pro doing industrial robot IK simulation in one vibe-coding session is wild. The crossover between agentic reasoning and physics sims is exactly where real-world autonomy gets built. this level of integration changes the game for hardware + AI teams

Gemini 3.1 Pro improves a lot at the robotics level.

very cute style

thanks

Welp you earned a follow

awesome!

three.js ?

yes

amazing.

thanks

as always !! amazing 🔥

thanks Vamsi

😯

Thats bad ass!

thanks

Nano Banana is getting a lot of hype lately, I can see why.

@LucienReconq cool

Wow

As a licensed FANUC robotics operator, I approve of this message

super cool, especially having everything from design to simulation in one flow, feels like a new kind of builder stack emerging here, would be great to connect

the IK simulation is where this gets really interesting. most vibe coding projects are web apps and dashboards... seeing someone push it into robotics with actual inverse kinematics is a completely different challenge. the spatial reasoning improvement in 3.1 makes sense too, IK solvers need the model to actually understand joint constraints in 3D space not just generate code that compiles

seeing this at the top of my feed gives me hop the algo is getting back on track

I feel like the next step is to take these industrial robots out of industries. We need to see them in stores, malls, and be more indulged in our daily life.

Could you please show me the command for this simulation?

This is wild! Gemini 3.1 Pro crushing the IK simulation. The key insight: for physics-heavy tasks, Gemini's spatial reasoning shines. For larger codebases, Opus still wins. Pro tip: use noise functions for metrics that don't need real data - saves compute and looks legit. 🔥

seems like an inefficient way to move boxes 1 foot

This is exactly where AI gets interesting — when you chain models together. Nano Banana for design. Gemini Pro for logic. Next step: add video generation for the full simulation. The future of robotics dev is multi-model pipelines, not single-model magic.