Arthur Allshire

Excited to introduce QF3: Fast Flow RL with Filtered Q-Gradients!

Train humanoids from scratch, fine-tune VLAs, even fine-tune image models... all with one simple off-policy update ✨

Humans generalize from rich interaction experience. We bring this to humanoid loco-manipulation through scalable real-to-sim-to-real.

Introducing PRISM, we expand 4 real videos into 256 counterfactual variants, recover robot-object trajectories, and train one policy that generalizes across diverse objects, scales, and spatial layouts.

CoRL’26 🚀 Website for details 👉 prism-real2sim2real.github.io

Work done at Amazon FAR w/ @zhenkirito123 @pabbeel @rocky_duan @JitendraMalikCV @carlo_sferrazza @ckarenliu @GuanyaShi @akanazawa

Almost all bimanual robots include wrist cameras. They add bulk, cost, and complexity, but are critical for performance. Are they really needed?

Excited to finally share EyeRobot 2.0, where we use active gaze from an ego stereo camera instead 🧵 (with @KushtimusPrime)

check out our sim and code release! some super useful envs for evals, rl etc for people to try out!

On overfitting being helpful, I don’t think it’s unique to robotic BC
eg. from InstructGPT “we find that our SFT models overfit on validation loss after 1 epoch; however, we find that training for more epochs helps both the RM score and human preference ratings”.
I think the important distinction is pre-training vs post-training. Val loss is a useful metric for the former but not for the latter

Hand teleop is an import yet challenging problem, Adam / Sarthak's project is a really great step towards making such systems accessible and reliable, check it out!

really excited to have got this out!

here's an uncut, real-time video of our policy folding a cardboard box

if you take GPU maxis at their word, they’re the swiss army knife of computing

but if they can be easily slotted into any training job, their deployments have no stickiness bc a competitor can also easily replace them

they have to monopoly win or margins go to zero

bearish...

super cool work from @tylerlum23! some of the most dexterous hand-arm manipulation I've seen yet

slides look good!

waiting for someone to open an sf cocktail bar called expert of mixtures

one of the best moments at BAIR lab actually never imagined to spot prof @redstone_hong on a random day

really excited about this work, finally enabling learning locomotion end to end from RGB pixels!

super excited that we won the best student paper for videomimic!

unfortunately i was packing in the hotel during the awards ceremony 😅

cool work. recommend digging into details -- they show how throughput can be increased by changing the arrangement of envs and RL training across a node. this will become especially important as policy size increases with pre-trained models..

excited that VideoMimic was just nominated for an award at CoRL, see you soon in Seoul! 🇰🇷

we have also now also released our sim/real-robot deployment pipeline, including checkpoints and nice viser web viz -- check it out
github.com/hongsukchoi/Vi…

viser has been the biggest quality of life improvement for me in years for visualising robotic data & sim / debugging policies. congrats to brent on the release and check it out!

training diffusion-style models with RL has traditionally required heavy re-engineering of the MDP. david’s paper removes the need for this and provides a clean way to train diffusion with RL using a tweaked PPO style objective. super exited by this — check it out!

really cool project from @tylerlum23 on learning manip from human videos! some pretty impressive and dexterous tasks in here!

check out pyroki! we used it in videomimic and it's super useful to be able to do things beyond normal retargeting for high-DoF systems (eg contact aware was critical).