Hongsuk Benjamin Choi (@redstone_hong) — posts, photos & videos

We introduce FlashDexRetarget: one RL policy that retargets many human hand-object demos to dexterous robot hands.

On a 50-motion benchmark with XHand: 90% success with ~100x less compute than CHORD.

Project page: holiday-robot.github.io/FlashDexRetarg…
Paper: arxiv.org/pdf/2610.01849

2:59

12 21 7 189 16.7K 155

@redstone_hong reposted

Amil Dravid @ COLM2026

@_AmilDravid

Check out this work co-led by @sophielwang. We find that base models can match much of RL’s reasoning gains with the right first tokens, which we call 𝘵𝘰𝘬𝘦𝘯 𝘤𝘶𝘦𝘴. The effective cue differs across models and can be as simple as “.\n\nOkay,” or “To determine,” with no explicit instruction to reason. We dissect this phenomenon through a set of interventions on RL and the base-model training data itself. We also extend our analysis to a safety case study of refusal and compliance.

Token cues can arise from simple associations in reasoning-related training data, and we can even turn the word “chicken” into a reasoning cue through counterfactual mid-training runs.

Super excited to find a rich underlying phenomenology! Check out Sophie’s thread for the main findings. I’ll share a few other observations below that I found especially cool. 👇

Sophie Wang@SophieLWang · Oct 6

Can “chicken” make a base model reason? 🐔

Yes!

Why? With the right first tokens, base models can match RL in reasoning. We trace this to learned training data associations and use a simple data edit to make “chicken” elicit reasoning too.

blog: sophielwang.com/cues

0:11

4 19 1 72 21.5K 22

@redstone_hong reposted

Carlo Sferrazza

@carlo_sferrazza

Turns out the FastTD3 recipe transfers really well to flow policies, and you keep all their benefits. Check out QF3, the new algorithm @ChungMinKim developed during her internship at FAR!

Chung Min Kim@ChungMinKim · Oct 6

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 ✨

1:05

3 4 0 53 4.3K 21

@redstone_hong reposted

Chung Min Kim

@ChungMinKim

The same update also makes a pretrained VLA in ABC's simulation (abc.bot) faster and more reliable -- and fine-tunes full flow policies on Robomimic too!

0:35

1 1 0 24 1.1K 4

@redstone_hong

Oct 7

Value-based RL matters for robot policies because we can't condition on everything, at training time or at inference time. Conditioning picks from what's in the data; a value function can find something better. QF3 shows efficient behaviors emerging in manipulation from simple off-policy update. Check it out!

Chung Min Kim@ChungMinKim · Oct 6

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 ✨

1:05

0 3 0 36 3.1K 24

@redstone_hong reposted

Seohong Park

@seohong_park

Doing proper science in robotics is hard, because human demonstrations are expensive and less controllable.

OCBench is a new benchmark providing "human-like" scripted policies. This enables scientific studies of BC and RL with virtually unlimited data!

0:46

29 63 8 572 29.6K 299

@redstone_hong reposted

Kush Hari

@KushtimusPrime

Oct 2

Excited to release EyeRobot 2.0! We noticed it is better at dealing with distractors. When the gaze policy looks at the right target, the policy can recover

0:34

Justin Kerr@justkerrding · Oct 2

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)

0:34

3 10 1 90 6.8K 36

@redstone_hong

Sep 30

impressive work!

Zihan Wang@Z1hanW · Sep 30

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

1:00

1 0 0 15 2.5K 2

Nicolas Keller@Nicolas_Keller · Sep 29

@redstone_hong

Sep 29

amazing! where are the humans resetting the scenes before rollouts?

The Era of Evals is coming to AI robotics.

Today we’re launching Reality Check: our leaderboard & first public robot manipulation benchmark.

14,400 real-world rollouts. Four models. Multiple tasks & data regimes. Object placements. Confidence intervals. FR3 Duo stations.

1/10

0:23

1 0 0 28 3.9K 5

@redstone_hong

Sep 24

super light! amazing

UploadVR@UploadVR · Sep 23

Meta VR Glasses were just officially announced. They have micro-OLED displays, weigh 100 grams, and are powered by a tethered puck with Qualcomm's Snapdragon Reality Elite chipset.

Details here: uploadvr.com/meta-vr-glasse…

1:10

1 0 0 9 1.6K 1

@redstone_hong

Sep 18

Love to see robotics research paying attention to real challenges!

Chelsea Finn@chelseabfinn · Sep 15

A video from a Pi robot deployed at Dandelion Chocolate, fully autonomous w/ no interventions. 🤖

Deploying robots has taught us surprising lessons about the gap between proof-of-concept (i.e. building one box) and real-world utility (productively building boxes for hours).

I expected that the hard part is building the box, since it’s the most dexterous, but that wasn't what we found. Counterintuitively, the hardest part was reliably stacking the boxes. Our original table-mounted robot had poor visibility of the stack without special separately-mounted cameras. Plus, stacking requires more generalization (each box is placed in a different location), and an imprecisely-placed box can lead the entire stack to collapse many boxes later.

We recently switched this deployment over to a mobile robot, and it's fun to watch the robot being completely self sufficient for multiple hours. 🙂 Data and feedback from real world use-cases like these are quite valuable for the π pre-trained model as we scale!

4:19

0 0 0 11 1.4K 1

@redstone_hong

Sep 18

ABC is fully out and headed to CoRL 2026!

ABC-130K is the largest open teleop dataset to date: 3,500 hours, 130K+ episodes, 195 tasks, collected on an $8K bimanual setup. The full stack is open: hardware, training code, sim, and eval, so academic labs can do this research on equal footing.

Beyond the data, we ran the baseline science for the community to stack on:
• Sim-to-real correlation (r = 0.91 on task progress)
• Which offline metrics matter?
• Scaling law in compute (training steps and batchsize)
• Conditioning the policy

Project page: abc.bot
Github: github.com/amazon-far/abc

Adam Rashid@adamrasb · Sep 18

Happy to announce the full ABC release! We’re also excited that ABC was accepted to CoRL 2026!

Check out our website for code, 400+ hours of sim data on 24 tasks, and 5,850 labeled policy-evaluation episodes.

@arthurallshire @Cinnabar233 @ritvik_singh9 @redstone_hong @davidrmcall

1:30

1 10 1 89 25K 55

Hannes von Essen@HannesVonEssen · Sep 7

@redstone_hong

Sep 7

cute! we need some cute robots

🦆
🏀

0:20

0 0 0 3 791 0

@redstone_hong

Sep 3

interesting perspective

Zhen@zhen_do_ob · Sep 2

So uh, the winners that will likely emerge will likely be Dyna and Skild going head to head:

1) PI is cooked because they don’t deploy well, don’t have good research techniques and the talent flow there is not S tier - no wonder they were evaluating selling to Ant
2) Generalist is gonna be slower because they need to partner with startups to deploy and do evals which means their speed and knowledge won’t compound as quickly since the deployment partners have competing priorities
3) Genesis is just getting to lab demos and just starting on hardware so they’re behind
4) Sunday is full stack but they also haven’t truly deployed into real consumer environments and is competing vs Figure - the 99.1% solve of folding was ok but the demos clearly show they can’t get tput up.

So yeah, it’s pretty clear to me this will be a Dyna vs Skild two horse race since both have deployed, both have real revenue, but afaik only Dyna has successfully grinded down the 9s on quality, throughput, and cost for multiple real world customers while having all of the speed, knowhow, etc to compound both the data flywheel + b2b network effects.

QED I rest my case.

1 0 0 7 2.9K 0

@redstone_hong

Aug 27

We saw exactly this in the ABC paper (abc.bot) on real robots. Across 16 checkpoints, validation flow loss was uncorrelated with real-world success, while validation action error had the strongest correlation. Fig 8, if you're curious.

Seohong Park@seohong_park · Aug 24

Behavioral cloning mystery

seohong.me/blog/behaviora…

I wrote a new blog post about "mysteries" in behavioral cloning that appear with real-world robot data (e.g., overfitting is "good"). I also tried to demystify them and shared my thoughts!

2 5 0 96 11.8K 79

@redstone_hong

Aug 14

very cool!!! i love this work

Adam Rashid@adamrasb · Aug 14

Excited to share SPD: simulation pre-training for dexterity. We pre-trained a policy in simulation and fine-tuned with less than 2 hours of real data (with @sarthakkamat)

1:14

0 0 0 14 2K 2

@redstone_hong

Aug 9

agree with Zhi-Yi

Eren Chen@ErenChenAI · Aug 9

VLA is dead?

Zhi-Yi Ren, a Research Scientist at Physical Intelligence⁠ working on VLA pre-training, argues that VLA and World Models aren’t competing paths.

His view: modeling the future can simply become part of the VLA stack and the two approaches may eventually converge.

0:46