Yanda

robots @uwcse, prev @Roblox

[Yanda @BaoYanda]

@PrajwalAvhad8 figure backers made this chart

[Yanda @BaoYanda]

Sequential data is awesome because ICL is key to generalization, as shown by LLMs. I don't think demonstration is the best way to generate that sequential data though - you get few shot capabilities like generalist's demos but not true ICL. For that you would need traces of teleopers improving at the task. But then the history is not on-policy, so not as effective. So DAgger seems like the way to go, but distillation from DAgger is only a one-step improvement operator. There has to be a better way...

[Siqiao Huang @KnightNemo_]

I would end by a couple closing remark:
1. In-context learning for robotics is a long-standing problem;
2. Scaling context length (& format diversity) is a commonly-suggested advise for robotics. However, does/should this icl/reasoning capability happen at the VLA layer rather than the VLM layer? Since VLA layer inference needs to be served at a certain frequency for closed-loop control, and naive attention's compute scale quadratically with context length;
3. If you combine this half-a-year+ training with DYNA-2's curve, the takeaway should not be let's scale to 1B hours of robot data & 1 million gpus for compute, but there's some fundamentally wrong with the current paradigm, and there's something crucial missing in the robot learning repertoire.
What an exciting time to be alive.
6/6

[Yanda @BaoYanda]

@ericcouu Bro just wanted to post that he had a 100m views

[Yanda @BaoYanda]

@ahmxrd MoEs share weights, like at attention layers. It's also just better to train at scale across domains because there's transfer that improves capabilities as a whole

[Cornelius Braun @corbraun]

𝗚𝗼𝗼𝗱 𝗺𝗮𝗻𝗶𝗽𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝗺𝗮𝘆 𝘀𝘁𝗮𝗿𝘁 𝘄𝗶𝘁𝗵 𝗯𝗲𝘁𝘁𝗲𝗿 𝘀𝘁𝗮𝘁𝗲𝘀, 𝗻𝗼𝘁 𝗯𝗲𝘁𝘁𝗲𝗿 𝗿𝗲𝘄𝗮𝗿𝗱𝘀.

We sample diverse, physically feasible contact states as starts + goals for RL.

The behaviors that emerge are surprisingly dynamic and reactive: including recovery, regrasping, and adapting through contact. 👇 (1/n)

[Yanda @BaoYanda]

@andrewho03 Is this because model capability is extremely spiky so not good outside coding or lack of integration w clients, in which case long lab consulting arms? If application layer is still just using frontier models as base then seems to be the second

[Yanda @BaoYanda]

@davehpitt A good enough layer consumes the vertical, as seen in LLMs

[Yanda @BaoYanda]

@stisov16 This is like the 10th paper to claim a scaling law, and not just on WAMs

[Yanda @BaoYanda]

@jamesw_stevens Congrats on the launch!! Excited to see what DAgger at scale can do

[Yanda @BaoYanda]

In my experience, claude is much better at long horizon execution, while Gpt is significantly better at teaching.

So basically claude is maradona and gpt is scaloni

[Yanda @BaoYanda]

@Skrillex come to the CS building at the university of washington skrillex! only a few blocks down :D

[Chuning Zhu @chuning_zhu]

Can we get robots to think before they act? In Latent Memory Palace, we achieve this through a classic principle: variational inference. (1/8)

[Matthew Hong @matthewh6_]

RL can't teach an LLM to solve problems it never solves. Robot policies suffer from exactly the same limitation.

The fix turned out to be almost embarrassingly simple: train the policy with diffusion noise during pre-training.

That's it.

The policy covers a much wider action distribution, RL finally has somewhere to search, and we fine-tune VLAs on real robots in under an hour.

Introducing TMRL.

🧵(1/9)

[Yanda @BaoYanda]

@CharlieZZhang1 Boyfriends are gonna face a spike in job loss

[Yanda @BaoYanda]

@pranavcmadhukar the issue is that pretrainings prior is indispensable

[Yanda @BaoYanda]

is it not optimized for implicitly (which is enough)? internet scale next token prediction across many diverse documents is impossible w/ out doing some meta learning right? When training on a blog, it can be written by either a professor or a high schooler. only a few samples/tokens from the current task will help you infer who it is/lower training loss. So when pre-training, meta learning has to be done

[Tyler Westenbroek @ty_westenbroek]

Real-world RL is still too brittle and data-hungry for long-horizon, contact-rich tasks.

We introduce Simulation Distillation (SimDist), which turns large-scale simulated experience into reusable world-model priors for rapid real-world adaptation.

By combining online planning with dynamics adaptation, SimDist achieves high success rates on tasks requiring precision, force, and reactivity.

Play with our interactive visualization to see for yourself: https://t.co/qFGNySxdAl

(1/n)

[Abhishek Gupta @abhishekunique7]

Punchline: distill world models from simulation to enable fast, stable real-world robot adaptation.

Simulation is nearly always wrong. But in Simulation Distillation, we ask a simple question:

How do we perform simulation pretraining such that real-world adaptation becomes trivially easy?

https://t.co/ORDaxU2gzs

Let's take a closer look (1/n)

[Reuben Santoso @reuben_santoso]

Imagine testing your app like real users actually use it. not one fake user clicking through a script, but many users interacting in a single workflow.

One uploads a post. Others like and comment on it. In the same test run. In real browsers.

@JoveW, @SamuelJepee and I built @qualtydotco a agentic QA that evolves test cases with your product and finally catch the hardest bugs that only show up when users interact. No brittle scripts. No Maintenance.