Ammar

robots for the future @purposerobotics ML/AI @Autodesk | Private Pilot | @fdotinc | Tron Eng

[Ammar @aamarsbarr]

@nickventuri @interaction Poke be roasting me after some of the meals I eat/log 💀

[Ammar @aamarsbarr]

Just built a fitness app for poke.

✅ track calories with just a simple text or picture of your food
✅ have poke coach you and be ur personal workout buddy

Dm me if u want me to send u the beta link.

[Poke @interaction]

Say hi to the new Poke! 🌴

Now officially approved by Apple to text on Apple Messages.

As the first and only AI agent. Chat now: https://t.co/VIWYU64dUI

[Ammar @aamarsbarr]

@MilHoornaert @interaction It’s just convenience. There tons of other ways out there like spreadsheet but I find using an LLM/agent to be lot easier and natural.

[Ammar @aamarsbarr]

@conorsoup @interaction DMs should be open now.

[Ammar @aamarsbarr]

@raulg12376 @interaction Just dmd

[Ammar @aamarsbarr]

@beardthelyon @interaction Oh yea it does. Absolute game changer. it pulls steps and active calories from apple health. Hoping to try and integrate more soon.

[Ammar @aamarsbarr]

claude fable 5 just dropped....a mythos class model. rip my api bill

[Ammar @aamarsbarr]

Anyone at ttw homecoming wanna chat robots?

[趣味開発をするアライグマ @arai01robo]

Robotics Engineer in 2026….

[Ammar @aamarsbarr]

Just stumbled upon a new paper called COBALT.

COBALT = = crowdsourced robot teleop from smartphones

While the control interface of teleop with smartphones is interesting the paper goes deeper in exploring a scalable cloud infrastructure for collecting imitation learning data and remote teleop

Maybe this is the universal way to bridge the gap?

Give it a read: https://t.co/B6xiudKqJ4

[Ammar @aamarsbarr]

@whosjunaidd People like junaid are the reason why startup culture is awesome!

Shoot me a dm as well and I’ll try what ur building!

[Google DeepMind @GoogleDeepMind]

We’re dropping Gemini Omni: our first step towards a model that can create anything from anything - starting with video.

It combines Gemini’s intelligence with our generative media systems - representing a leap forward in world understanding, multimodality, and editing 🧵

[Ammar @aamarsbarr]

For new upcoming research ArXiv is ur best friend.

Just found this super interesting one about how they trying to detect “false claims” of returns for Amazon, DoorDash and Uber Eats by seeing if it’s an AI generated photo or a real one.

[Ammar @aamarsbarr]

One of the biggest bottlenecks in VLA models is long horizon reasoning.

Standard policies just look at the current frame -
π(aₜ | imageₜ, instruction) and this works fine for short tasks, but after 20–100 steps, minor errors compound, the robot drifts and everything breaks down.

TraceVLA proposes giving the robot an explicit visual memory of its past trajectory without bloating the context window.

Instead of passing a massive history of raw frames, it overlays motion traces directly onto the current image using a dense point tracker

The pipeline is super simple:
1.Track keypoint trajectories from past frames
2.Filter for actual movement
3.Overlay those trajectories onto the current observation frame
4.Feed both the clean image and the trace image into the VLA backbone

Input sequence: [Image tokens] → [Trace tokens] → [Language tokens] → Transformer → Action

Instead of forcing the transformer to reason across a long temporal context window, it literally compresses trajectory history into a spatial representation. The robot doesn’t need to remember the entire action sequence because it can just see the path it already took.

The results speak for themselves:
• +10% improvement over OpenVLA in simulation
• 3.5× higher success rate on real robot tasks
• Strong cross-embodiment generalization

Check the OG paper: https://t.co/Ylw5PTqAcZ

[Ammar @aamarsbarr]

Interesting blog post about humanoid data. Which type do yall use?

https://t.co/sWSleSHHJ8

[Ammar @aamarsbarr]

Bro released the algo 😳

[Elon Musk @elonmusk]

The latest 𝕏 algorithm has been published to GitHub
https://t.co/ZCOm51uxmh

[Ammar @aamarsbarr]

What if robots could imagine…

As I was learning more about VLAs I came across smth called VAMs

The core idea is that just like humans can “imagine” what’s going to happen and “visualize” the action for picking up a cup or placing it in a box VAMs aims to do the same with robots

And in my head the same way u train a policy by teleportation and mimicking the movement, this is tryna recreate the same concept.

This was the one I found: https://t.co/b0lLHrolWP

Lmk ur thoughts, if I misunderstood smth or if u wanna chat about it in more detail.

[Ammar @aamarsbarr]

Just built an app that lets me check and work on my projects via codex, claude or gemini from the gym. Absolutely life changing.