We are releasing Dyna-2.1, the first Physical Agent that achieves reliable super long-horizon whole-body autonomy. It combines our brand-new semi-humanoid hardware with an agentic system built around Dyna-2 to handle ultra-long real-world workflows.

Here is an uncut footage of entire hour-long laundry room workflow, just like a human does.

After so many demos, models, pilots, robots are still struggling to land real deployments with real customers. Until now.

Today we’re excited to share that our robots have successfully crossed the ROI threshold, and Din Tai Fung, one of the highest revenue per location US, is rolling out Dyna robots across its extensive restaurant network.

This brings our rollouts across hotels, logistics, data centers, and many other use cases to a fleet that reaches hundreds of robots by the first half of 2027. And we’re just getting started.

It’s been a wild year, and today we’re double clicking on the battlefield stories and sharing a few learnings about scaling robot deployments. We are just scratching the surface.

Read the full blog post: dyna.co/news/scaling-c…

When people talk about robotics, they usually talk about models, data, or hardware. Few people talk about the infrastructure that lets you iterate on all three quickly. Today we're publishing how we trained Dyna-2 on over 1,000,000 hours of egocentric video, repeatably. At this ked at ten thousand hours did not hold up:

• ingestion throughput was capped at 14,000 episode-hours per week — a million hours would have taken over a year
• building a training manifest took 48 hours before a run could even start
• reading a petabyte from cloud storage during training left GPUs exposed to latency and packet loss

🧵

🚀New Paper
arxiv.org/abs/2606.28551

Everyone obsesses over VLM architectures & training recipes. But what about the data?

Presenting the latest work in the DataComp-series: a testbed for VLM data curation with 1,000+ controlled experiments and some surprising lessons 👀

🧵👇

1/ We fine-tune a lot of customer models, so we decided to systematically try and figure out some best practices for finetuning. SFT isn't sexy, but it's still important. We vary one SFT lever at a time across 2 model families, dense + MoE to 235B, on 4 real-world customer his clean is that each dataset is paired with an eval that took weeks to build with the customer, and the training outputs were generated to pass that eval. So the supervised target and the thing we measure downstream are the same criterion, which strips out the usual confounders

My dream robot task! Teach it to do passing with a human!

Our white paper just came out on Arxiv arxiv.org/abs/2604.19728…. We open-sourced it all github.com/TRI-ML/vla_fou…. Our project website also has links to the white paper, the weights, and more videos tri-ml.github.io/vla_foundry

In LBM Eval, our Drake-based multi-task tabletop manipulation simulation:
• Our from-scratch VLA matches our prior LBM
• Our VLA using a Qwen3-VL backbone tops that by over +20%

Shout out to the amazing team that made this possible: @sedrickkeh2 @karora4u @yoloswag @imp_aa @s1wase @robo_kat I’ve enjoyed building this with you. We’re all excited to see what the community builds with it!

Releasing VLA Foundry: an open-source framework that unifies LLM, VLM, and VLA training in a single codebase. End-to-end control from language pretraining to action-expert fine-tuning — no more stitching together incompatible repos.

The “large behavioural models” built by @ToyotaResearch for @BostonDynamics Atlas is available in @LeRobotHF

LBMs are under-explored compared to VLAs and video/world action models, I’d like to give it a go at some point 😄

Thank you @brysonkjones for open-sourcing!

Today I’m releasing an open-source implementation of Multitask Diffusion Policy model, most widely seen deployed on the Boston Dynamics Atlas Humanoid Robot!

Check out the demo + some cool takeaways

I’m so tired of writing rebuttals to this kind of “lack of novelty” review: “This paper trivially combines A, B, and C, so the algorithmic novelty is limited.”

Technically, most (if not all) robotics papers are convex combinations of existing ideas.

I still deeply appreciate A+B+C papers—especially when they deliver:

- New capabilities: the “trivial combination” unlocks behaviors we simply couldn’t achieve before

- Sensible & organic design: A+B+C is clearly the right composition—not some arbitrary A′+B+C′

- Nontrivial interactions: careful analysis of the dynamics, coupling, or failure modes between A, B, C

- Rehabilitating old ideas: A was dismissed for years, but paired with modern B/C, it suddenly works—and teaches us why

- System-level & "interface" insight: the contribution is not any single piece, but how the pieces talk to each other

- Scaling laws or regimes: identifying when/why A+B+C works (and when it doesn’t)

- Engineering clarity: making something actually work robustly in the real world is not “trivial”

- New problem formulations: sometimes the real novelty is in the reformulation—only under this view does A+B+C make sense.

Maybe worth keeping these in mind when reviewing the next A+B+C paper : )

Baking without premix.

@sedrickkeh2 @robo_kat @karora4u @s1wase @yoloswag @ToyotaResearch

𝑪𝒐-𝒕𝒓𝒂𝒊𝒏𝒊𝒏𝒈 is a promising way to scale Large Behavior Models (LBMs) beyond robot data, yet the data and training recipe are far from settled. 🤔

We present a large-scale empirical study leveraging 4,000h of robot/human data and 50M vision-language samples, evaluating 00 simulation rollouts and 2,835 real-world trials. 🤖📊

co-training-lbm.github.io
Work done during my internship at @ToyotaResearch.

How can we make a better TerminalBench agent?
Today, we are announcing the OpenThoughts-Agent project.
OpenThoughts-Agent v1 is the first TerminalBench agent trained on fully open curated SFT and RL environments.
OpenThinker-Agent-v1 is the strongest model of its size on a new bar on our newly released OpenThoughts-TB-Dev benchmark. (1/n)

📢📢📢 Releasing OpenThinker3-1.5B, the top-performing SFT-only model at the 1B scale! 🚀

OpenThinker3-1.5B is a smaller version of our previous 7B model, trained on the same OpenThoughts3-1.2M dataset.

The short version is: LBMs work! We see consistent and statistically significant improvements as we increase the amount of pretraining data. But doing the science is still hard; as a field we have more work to do to improve the statistical power of our experiments.