Ben ✈️ CoRL 2026

Research Scientist @LumaLabsAI. Core contributor to Uni-1. Now doing RL for robotics

[Ben ✈️ CoRL 2026 @bentropy_ai]

(to clarify - we recreated the tasks from their release video)

[Ben ✈️ CoRL 2026 @bentropy_ai]

Fable 5.2's launch impressed with strong robot manipulation. How does ZAP, our 27B model, do on the same tasks?

We recreated the tasks in sim and ran both through the same harness (the model is the policy, calling robot tools directly). ZAP only trained on sim pick-and-place, so every object and goal is new to it.

Task 1, open a bottle: ZAP is 2.5x faster and costs $0.43 vs. Fable's $6.50. It plans the grasp up front, and it pays off: a firm hold and clean turns. Fable drops the bottle, recovers nicely, and still finishes.

Sim only for now since I'm at CoRL, but the real robot is waiting back home 👀 🧵

1/4

[Ben ✈️ CoRL 2026 @bentropy_ai]

Astra made robot control look like a job for frontier LLMs.

But small VLMs already know what to do. They just don't know how. Turns out, simulation is all you need.

Meet ZAP. No real robot data, no frontier teacher. Matches Astra on real tasks, 3× faster, 17× cheaper. 🧵
(1/11)

[Ben ✈️ CoRL 2026 @bentropy_ai]

The last task in their demo vid, solving a Rubik's cube, is still on the to-do list (I need to build the sim cube first 🧩).

Scorecard: both models finished all 3 tasks. ZAP was 1.3–2.6x faster and 7–15x cheaper, but Fable was cleaner on 2 of 3.

Still plenty of work left for agentic robots on precision, cost, and speed. More soon!

https://t.co/zOIW0E6gjn

4/4

[Ben ✈️ CoRL 2026 @bentropy_ai]

Task 3, assemble a gear set: fit a gear into place so it meshes with the others, then turn the center gear. This one demands precision, and Fable shines with cleaner grips and more precise placement.

ZAP's placement is less tidy, but it's nearly 2x faster, takes fewer turns, and costs $0.24 vs. $2.54.

3/4

[Ben ✈️ CoRL 2026 @bentropy_ai]

Come say hi! Presenting ZAP at 12-1:30 at Griffin Hall, poster 134

[Ben ✈️ CoRL 2026 @bentropy_ai]

Who says robot control needs a giant model? ZAP, a small VLM, matches Astra on real robot tasks - 3× faster and 17× cheaper.

Presenting tomorrow at #CoRL2026. Come chat!

📍 Griffin Hall · Poster 134
🗓️ Mon, Nov. 9 · 12:00–1:30 pm

[Ben ✈️ CoRL 2026 @bentropy_ai]

@DominiqueCAPaul Less than you’d expect! We spent a lot of time figuring out how to optimize the 3 different compute sinks during RL - training, inference rollouts, and simulation (led by @djsaunde @mirceamironenco @DoubleMathew @thomasneff93 and Salman). Full details to come soon!

[Ben ✈️ CoRL 2026 @bentropy_ai]

Astra made robot control look like a job for frontier LLMs.

But small VLMs already know what to do. They just don't know how. Turns out, simulation is all you need.

Meet ZAP. No real robot data, no frontier teacher. Matches Astra on real tasks, 3× faster, 17× cheaper. 🧵
(1/11)

[Ben ✈️ CoRL 2026 @bentropy_ai]

Who says robot control needs a giant model? ZAP, a small VLM, matches Astra on real robot tasks - 3× faster and 17× cheaper.

Presenting tomorrow at #CoRL2026. Come chat!

📍 Griffin Hall · Poster 134
🗓️ Mon, Nov. 9 · 12:00–1:30 pm

[Ben ✈️ CoRL 2026 @bentropy_ai]

Astra made robot control look like a job for frontier LLMs.

But small VLMs already know what to do. They just don't know how. Turns out, simulation is all you need.

Meet ZAP. No real robot data, no frontier teacher. Matches Astra on real tasks, 3× faster, 17× cheaper. 🧵
(1/11)

[Ben ✈️ CoRL 2026 @bentropy_ai]

The awesome team behind ZAP at @LumaLabsAI: @djsaunde, @mirceamironenco, Salman Mohammadi, @doublemathew, @srzhu97, @_DivyaMakkar, @xuhuahuang01, @qinboli, @nsupreeth5, @censiyuan1, @Ir1dXD, Shyamal Buch and @thomasneff93

(11/11)

[Ben ✈️ CoRL 2026 @bentropy_ai]

You don't need a frontier model to control a robot. An open VLM already knows what to do; simulation can teach it how.

Learn more at https://t.co/1oPYvS8YYG

(10/11)

[Thomas Neff @thomasneff93]

Very excited about this! We've been cooking internally for to try and build a very fast & exact sequence parallelism strategy. TL;DR: it's really hard to find the one strategy to rule them all, but based on thorough evaluations and kernel work, we found a very strong overall approach :)

We'll be at PyTorch conference tomorrow and wednesday - if you want to learn more, please reach out! Also code & a more detailed technical report on this will follow!

https://t.co/k7mIdo9TF9