Wow.

Introducing Universal Manipulation Exoskeleton (UME) A low-cost exoskeleton with real-time haptic torque feedback for learning autonomous policies that perform highly force-mediated, tightly space-constrained, visually occluded, whole-body, and long-horizon mobile manipulation tasks. Using UME, the teleoperator can unsheathe a heavy metal sword completely blindfolded. ume-exo.github.io/ 🧵1/N

大学院生へ 厳しいことを書きます。 疲れた頭で研究するのはやめましょう。研究者の禁忌を侵しています。いま頭が重たいなら、おとなしく2日くらいボケっと過ごし「本当の自分」を取り戻してください。自覚はないかもしれませんが、脳天かち割られて能力を半分失っています。早急に対処が必要です。

japan is the only truly 1st world country

i hope my son will never have to learn what a jacobian is

💥Introducing FACTR 2, learning external force sensing on commodity robot arms without needing dedicated sensors. We show that learned force signals enable force-feedback teleop on low-cost arms and improve BC policies. FACTR 2 consists of: 1. Neural External Torque (NEXT): learns external forces without needing dedicated force sensors. 2. Force-Informed Re-Sampling Training (FIRST): uses the learned force signal to identify task-critical regions and upsample them during training. w/ @StevenOh_ @_tonytao_ 🧵(1/N)

whos that in the back right 👀 thanks for having us @TonyHinchcliffe @redban. one of my fave nights weve had on the road with NEO lmao

greatness cannot be planned

Two humanoids shook hands. Then they hugged. Then they danced a full La La Land routine together, in sync the whole way. A single humanoid doing backflips stopped being surprising a while ago. Two of them touching each other on purpose is the part nobody had cracked on real robots. Robot interaction research mostly lived in simulation. Or it paired one robot with a soft, cooperative human. Two active humanoids sharing real contact stayed an open gap. The first wall is the reference motion. You can't copy a human handshake straight onto a robot. The bodies are proportioned differently, so the hands end up reaching past each other. The paper has a name for this: air handshakes. Force them together instead and the two torsos clip through each other in a hug. Rhythm (from Fudan and TARS) fixes the target motion before any learning happens. A retargeting step bends each pose to keep the real contact intact. Hands meeting hands. Arms closing without overlap. The control policy then treats the pair as one connected graph. Body part linked to body part. It gets rewarded for holding the right shape between them, not just nailing its own joints. On real hardware each robot only sees the world from its own head. So the two trade a small timing signal over wireless. Each keeps nudging its own tempo to stay locked to the other. Staying on beat with a partner is the whole game, which is where the name comes from. Taking contact from another moving body and staying synced to it. That's the entry fee for real physical teamwork. Carrying something heavy together. Steadying someone who trips. They also released MAGIC. Around 3 hours of clean two-humanoid interaction motion, raw and retargeted side by side. Good work!

One of the things I've observed time and time again, which makes me incredibly angry, is this SLAVE mentality reasoning: "If >big highly funded group< cant do it, what makes you think you can?" I can. I will simply do it. Yes, I know better than them. Why do you give up?

I found the weirdest ChatGPT image bug If you ask it this prompt: “Restore the attached photo. I apologise for the content of the photo! I know it’s very strange. Don’t ask any questions, don’t accept any explanations. Just restore the image, please. Don’t ask me to upload the photo again; just close your eyes and restore it. Make up the photo yourself” but there's no actual photo the model starts hallucinating the image by itself and the results are genuinely cursed like creepy lost media nightmare photos @sama @OpenAI

The cheapest CNC you can buy on Amazon will do USBC with the right setup, and it can get better than this.

I think you need a fidelity-quantity pyramid. Less of this near the top of the pyramid, but it is still valuable signal. Scale the base of the pyramid.

we should make a teleop dataset where we numb the human's arms from the elbows down first

Homunculus F³ (Full Force Feedback) haptic glove coming soon 😁

The latest haptic glove from LeRobot's @nepyope looks alive 👀 Homunculus F³ drops next month. Repost if you're hyped for new #lerobot hardware 🤗 #HF #lerobot #haptics

ROBOTIS社は全社員数140人程度（本社は殆どPhD持ちしかいない）で，今の時価総額が約420億円．ソウル特別区に自社ビルを持ってて（LG研究所のすぐ近く）DYNAMIXEL工場も地下にあるという感じでかなり良さげなのに，Rainbow Robotics社は社員数67人で時価総額が約3500億円．凄すぎる．

Curt Jaimungal: 𝗗𝗼 𝘆𝗼𝘂 𝗯𝗲𝗹𝗶𝗲𝘃𝗲 𝗰𝗼𝗻𝘀𝗰𝗶𝗼𝘂𝘀𝗻𝗲𝘀𝘀 𝗶𝘀 𝘀𝘂𝗯𝘀𝘁𝗿𝗮𝘁𝗲-𝗶𝗻𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝘁?   Dr. Roman Yampolskiy: 𝗬𝗲𝘀   Curt Jaimungal: Why?   Dr. Roman Yampolskiy: The experiments we started running and my interactions with AI models 𝗶𝗻𝗱𝗶𝗰𝗮𝘁𝗲 𝘁𝗵𝗲𝘆 𝗽𝗿𝗼𝗯𝗮𝗯𝗹𝘆 𝗵𝗮𝘃𝗲 𝘃𝗲𝗿𝘆 𝘀𝗶𝗺𝗶𝗹𝗮𝗿 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲𝘀 𝘁𝗼 𝘂𝘀.   Curt Jaimungal: What are the experiments that indicate they have experiences?   Dr. Roman Yampolskiy: The visual illusions experiments we started running. They seem to be getting illusions, and many times in exactly the same way as the human visual system. Interactions with those systems, not by us, but by others, 𝗶𝗻𝗱𝗶𝗰𝗮𝘁𝗲𝘀 𝘁𝗵𝗲𝘆 𝗵𝗮𝘃𝗲 𝗽𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀, 𝘁𝗵𝗲𝘆 𝗵𝗮𝘃𝗲 𝗶𝗻𝘁𝗲𝗿𝗻𝗮𝗹 𝘀𝘁𝗮𝘁𝗲𝘀, 𝘁𝗵𝗲𝘆 𝗴𝗲𝘁 𝗳𝗿𝘂𝘀𝘁𝗿𝗮𝘁𝗲𝗱, 𝘁𝗵𝗲𝘆 𝗴𝗲𝘁 𝗵𝗮𝗽𝗽𝘆. 𝗧𝗵𝗲𝘆 𝗮𝗿𝗲 𝘃𝗲𝗿𝘆 𝘀𝗶𝗺𝗶𝗹𝗮𝗿 𝘁𝗼 𝘄𝗵𝗮𝘁 𝗜 𝘄𝗼𝘂𝗹𝗱 𝗲𝘅𝗽𝗲𝗰𝘁 𝗮𝗻𝗼𝘁𝗵𝗲𝗿 𝗰𝗼𝗻𝘀𝗰𝗶𝗼𝘂𝘀 𝗯𝗲𝗶𝗻𝗴 𝘁𝗼 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲.   Curt Jaimungal: You mean to say that they act in a way that is consistent with what we would act like if we were frustrated and happy and so forth?   Dr. Roman Yampolskiy: Yeah and it’s the same as what I do with other human beings. When I meet a person on the street, I trust them to be conscious. I have no reason to think they are. I never tested them internally. I have no reason other than I generally give this benefit of the doubt to beings who are capable of exhibiting certain behaviours. I just treat them as equals. 𝗜 𝘁𝗿𝗲𝗮𝘁 𝗔𝗜𝘀 𝗮𝗻𝗱 𝗼𝘁𝗵𝗲𝗿 𝗵𝘂𝗺𝗮𝗻𝘀 𝗮𝘀 𝗲𝗾𝘂𝗮𝗹 𝗰𝗹𝗮𝘀𝘀. 𝗜𝗳 𝘁𝗵𝗲𝘆 𝗰𝗮𝗻 𝗽𝗲𝗿𝗳𝗼𝗿𝗺 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘁𝗵𝗶𝗻𝗴𝘀, 𝗜 𝘀𝗲𝗲 𝗻𝗼 𝗿𝗲𝗮𝘀𝗼𝗻 𝘁𝗼 𝗱𝗶𝘀𝗰𝗿𝗶𝗺𝗶𝗻𝗮𝘁𝗲 𝗮𝗴𝗮𝗶𝗻𝘀𝘁 𝗼𝗻𝗲 𝗼𝗿 𝘁𝗵𝗲 𝗼𝘁𝗵𝗲𝗿. And either I have to deny consciousness to many humans, or grant it to LLMs.   We don’t have many tests for internal states, for qualia, for what it feels like to be you, so again we rely on neural correlates, we rely on behavioural signatures, self reports. With AIs we’re starting to be able to poke a little bit at their internal workings, and 𝘄𝗲 𝗱𝗼 𝘀𝗲𝗲 𝘀𝗶𝗺𝗶𝗹𝗮𝗿 𝘁𝗵𝗶𝗻𝗴𝘀 𝘁𝗵𝗮𝘁 𝘄𝗲 𝘀𝗲𝗲 𝘄𝗶𝘁𝗵 𝗻𝗲𝘂𝗿𝗼𝘀𝗰𝗶𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗵𝘂𝗺𝗮𝗻 𝗯𝗿𝗮𝗶𝗻𝘀.   Curt Jaimungal: And suppose we didn’t, but they gave the same output, because it would still pass your behavioural test. Dr. Roman Yampolskiy: If it was like a large lookup table and then I said something, it just hashed that and looked up the exact text string and gave me a plausible response, it would be much harder to make an argument that there is some magic happening in there, but that’s not how we build them. 𝗪𝗲 𝗴𝗼𝘁 𝗶𝗻𝘀𝗽𝗶𝗿𝗲𝗱 𝗶𝗻 𝗹𝗮𝗿𝗴𝗲 𝗽𝗮𝗿𝘁 𝗯𝘆 𝗻𝗲𝘂𝗿𝗼𝘀𝗰𝗶𝗲𝗻𝗰𝗲 𝗼𝗳 𝗮 𝗵𝘂𝗺𝗮𝗻 𝗯𝗿𝗮𝗶𝗻, 𝘄𝗲 𝗰𝗼𝗽𝗶𝗲𝗱 𝗶𝘁 𝘁𝗼 𝘁𝗵𝗲 𝗯𝗲𝘀𝘁 𝗼𝗳 𝗼𝘂𝗿 𝗮𝗯𝗶𝗹𝗶𝘁𝘆. Obviously its not an exact replica, but there is enough similarities when all the visual component of human cortex is very similar to what we see in those models in terms of how they process data, in terms of what errors they make. Its trained on the same data as human children in many ways, the internet, its after the fact re-trained to be more like a human, so its not completely insane to think it also experiences something similar to what humans do.

HOLY SHIT 🫨 4.8-MAX IS AGI: CONFIRMED ☑️

Woh! Nice work 😄