
PhD student, researcher at @Mila_Quebec ~ Working on humanoids & RL
Canada
Joined September 2017
- Tweets934
- Following1.2K
- Followers239
- Likes3.2K
Alexandre Brown 🇨🇦 retweeted
(1/4) why do image models chop every picture into the same square patches? 🤔 (hint: they shouldn't)
following the bitter lesson, we let diffusion models *choose* how to tokenize an image. it spends fewer tokens on the boring background, and more tokens on our faces! 🧵
I was curious how much the style reward mattered in BFM-Zero.
So I removed it from the actor loss. The policy got significantly worse... and a lot more like C3PO 🤖, which I found quite amusing 😅.
Turns out style guidance matters more than I thought, even for steering the Forward Backward representations.
Motion tracking seems to hold up better without actor style guidance than goal reaching or zero-shot reward inference.
Still looks and performs slightly worse than with style guidance. Perhaps the constant C3PO stomping is just less awkward when the target is moving too!
1
1
53
Alexandre Brown 🇨🇦 retweeted
ChatGPT required RL, it was not pure imitation learning or dagger; reasoning models also required RL. Obviously the same will be true of robotics, the question is just if base models are good enough
Reproduced BFM-Zero results in sim. At 65M train steps, about 1/3 of the released checkpoint's 211M.
The policy already looks promising on these goal reaching examples. The released checkpoint appears stronger on some harder poses but the qualitative gap is smaller than I expected.
Zero-shot unsupervised RL looks promising 🔥!
More lookahead barely changed zero-shot motion tracking in my tests.
Averaging 1, 2, 4, 6, 8, 10, or 12 reference frames produced visually similar results for both checkpoints.
One possible reason: B receives reference velocities, so even a one-frame prompt (z) could encode motion information.
1
1
132
Zero-shot reward inference with BFM-Zero worked surprisingly well 👌.
My 65M steps vs the author's 211M. The flow is pretty simple: Define a reward function → infer z → act. No fine-tuning. The policy was never trained on any of these reward functions which is super cool IMO.
I observed strong behaviors overall, with room for more precise speeds and poses.
Random replay samples were used to compute z. No sample or prompt cherry-picking.
1
73
Alexandre Brown 🇨🇦 retweeted
GPT-6 Astra's robotic capabilities are not emergent. It was trained on robotics specific data. So were GPT-6.1 Sol and Gemini Argon
Here's how we know 🧵
Alexandre Brown 🇨🇦 retweeted
Replying to @theSamPadilla
It's the most open they have been in years ;).
Alexandre Brown 🇨🇦 retweeted
apple trained a unitree g1 humanoid only in simulation, across 1000 randomly generated houses
then it went into 6 real homes it had never seen, no fine tuning, no scans of the rooms
360 trials, 91% success
crouch under a table: 100%. go through a door: 93%. sit down: 85%
very bullish on sim
Alexandre Brown 🇨🇦 retweeted
the more i code with LLMs, the more i feel the urge to go back to fundamentals and understand things as deeply as possible.







