Johnny Núñez

Physical AI and Robotics at @NVIDIA

[Johnny Núñez @itsjohnnynunez]

@sama @OpenAI
My 93-year-old grandfather discovers ChatGPT Voice Mode for the first time, and the results are nothing short of amazing. He loved it, and it made him so happy. Affective computing and AI like this will transform life for older adults. #MerryChristmas

[Ming-Yu Liu @liu_mingyu]

NVIDIA Cosmos Lab is looking for PHD interns to help build Open Physical AI foundation models. The job is in Santa Clara, CA. Below is the application link.

https://t.co/vz21MpQEwQ

[NVIDIA Robotics @NVIDIARobotics]

NVIDIA Isaac ROS 5.0 brings AI agents to robotics development. 🤖

Announced at #ROSCon, the release adds agentic workflows, Isaac Skills, ROS 2 Lyrical support, GPU-accelerated libraries and NVIDIA Jetson support from Orin Nano to Thor.

Open source and available now 📖 https://t.co/6idfj67AZE

[Wall St Engine @wallstengine]

Nvidia $NVDA CEO Jensen Huang could pay around $8 billion in taxes over the next 5 years.

“I’m not afraid of paying taxes. I’m just afraid of being poor.”

“Instead of paying people as little as possible to get their work, I try to pay people as much as possible.”

[Johnny Núñez @itsjohnnynunez]

@matth_lapeyre So true!

[Seeed Studio @seeedstudio]

Move the target. The robot finds it again.

This visual grasping pipeline combines #IntelRealSenseD405 + #reBotArmB601-RS + #ROS2 + #YOLOE:

Detection → Depth → Hand-Eye Calibration → Grasp Pose → Pick & Place

Move the pen, and the system detects its new position, calculates a new grasp pose, and sends the arm to pick and place it.

The Seeed Wiki walks through the full #visualgrasping workflow — from camera & ROS setup to detection, calibration, grasp validation and the final grasp-and-place pipeline.

https://t.co/xHYzTqvJsC

[NVIDIA @nvidia]

The leather jacket returns to Washington, D.C. 👀

December 1, 2026.

🔗 https://t.co/K45B2DQ4kM

[Sadao Tokuyama @tokufxug]

sEMG×一人称カメラでロボット学習

🔹Wetour Roboticsが8ch sEMGリストバンド＋カメラを統合
🔹視覚で手の動き、sEMGで筋活動を取得
🔹視覚単体では手を21.8%のフレームで検出できず、最大4.32秒欠落

1秒分のsEMG処理は50.4ms。情報元はリプ欄。

[Seeed Studio @seeedstudio]

🤖 From natural language to physical action.

At #WRC 2026, we’re showcasing #reBot B601-RS × #VLM Skills—an agentic Physical AI system powered by #NVIDIA #Cosmos 3-Edge and 21 native Action Skills.

Perceive → Reason → Plan → Act → Observe → Recover

Moving beyond scripted #robot toward adaptive, closed-loop #PhysicalAI. 🦾

📍 WRC 2026 · Booth A435
📅 Aug 19–23 · Beijing

Read more: https://t.co/1fSzA2hsqC

[Yilun Du @du_yilun]

How can we improve robot generalization beyond collecting more demonstrations?

A robot can use inference to work out how to solve tasks it was never explicitly trained to perform. With learned models of the world, we can plan our future actions and goals before acting.

I wrote a research perspective, Generalization by Construction, on how learning can be combined with inference for flexible generalization.

https://t.co/MfNWubAp4l

[Johnny Núñez @itsjohnnynunez]

@ShuoYangAIR So true!

[Johnny Núñez @itsjohnnynunez]

For me, with the rise of agents, AI has become a trade-off between doing more extensive reasoning with fewer iterations, or doing less reasoning per step but relying on more iterations in the loop. That’s why I feel models need to become much faster and more efficient.

[Johnny Núñez @itsjohnnynunez]

Hard harnessing is the present and future of agentic robotics.

[Max For AI @MaxForAI]

来自普林斯顿的研究者 Yifan Zhang，今天发布了新一代循环回路Transformer架构：RLT（Recurrent Looped Transformer），相关技术报告已在 GitHub 公开。

它想探索的问题是：能不能让模型的内部计算持续延伸，不再被固定的网络深度限制？

先介绍一下作者。
此前，他参与过 NVIDIA Nemotron 团队的研究，也曾在字节跳动 Seed 团队担任 Top Seed 研究员。

有趣的是在发布帖里，他把这项工作与超级智能联系起来，甚至提出应当放缓前沿探索，直到安全的超级智能得以实现。

具体看方案，RLT 的核心，是让模型把上一轮的内部计算状态，直接交给下一轮继续处理。

它先用 Encoder 编码输入、建立上下文记忆，再交给循环运行的 Decoder。

每处理一个 token，Decoder 都会把最终隐藏状态送回下一个 token 的计算入口，同时保留各层的局部注意力缓存。

这套计算贯穿用户输入和模型回答，中途不会重置。

可以理解成，模型一边读、一边回答，内部还有一条不断接续的计算链。

上一步经过整个 Decoder 加工后的结果，会进入下一步，再经过一遍 Decoder。

这就是作者所说的无限推理深度：每个 token 执行的层数固定，但跨 token 串起来的计算路径，可以随着序列增长而继续延长。

例如，使用 48 层 Decoder，处理 100 个 token 后，沿着循环状态展开的路径就经过了 4800 个 Decoder 层。

它并没有真的变成一个拥有 4800 层独立参数的模型，而是在不同时间步反复使用同一套网络。

不过，目前公开的还是架构设计和技术报告，我看了下仓库里没有训练代码、推理实现或模型权重（应该很快就会开）。

作者也明确写了，推理能力提升、硬件效率和强化学习扩展效果，都还需要实验验证。

这个方向挺值得关注的。

最关键的问题是：模型能否真正利用这条更长的计算链，在同样的计算预算下，解决更多问题。

[Johnny Núñez @itsjohnnynunez]

@aronhouyu That is not a bar, is INS Land club

[WUJI @wuji_global]

MINT is now open source.

From first-person RGB video, MINT reconstructs 3D camera trajectories and bimanual hand motion in a shared world coordinate system. Reconstructed human hand motion can also be retargeted to WUJI Hand.

The release includes model weights, training and inference code, visualization and evaluation tools, EgoPipeline reference code, and structured annotations for 560,649 episodes covering 1,021 hours of first-person video.

Project:
https://t.co/9imtdqIHHH

Code:
https://t.co/NEyO1G9W5b

#MINT #EmbodiedAI #Robotics #WUJIHand

[Seeed Studio @seeedstudio]

#WRC 2026 Day 1 was a blast!🤖

At Booth A435, @seeedstudio × @RobStride_com are showcasing how #reBot B601-RS goes beyond scripted actions with #VLM Skills—turning natural-language goals into closed-loop Perceive → Reason → Plan → Act → Recover workflows, making reBot more agentic!

We’re also launching our new Physical AI course, collaborated with the #NVIDIA, taking developers from simulation to real-world #VLA deployment with reBot arm + NVIDIA #saac. Check full course at: https://t.co/COF5uchtI6

Huge thanks to the @NVIDIARobotics team for the continued support! 🙌 Come see Agentic #PhysicalAI in action at WRC 2026, Booth A435, Aug 19–23 in Beijing, China!

[OpenAI @OpenAI]

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.

The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.

The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

[Yunfei Xie @xiynfi1520580]

🤖 Introducing RoboTok, the “TikTok for robots.”

Just as TikTok recommends videos to people, RoboTok recommends relevant human demonstrations for robot learning.

Robot learning needs broad and diverse demonstrations, but collecting robot data is expensive. RoboTok is an internet-scale data engine that uses web video as a scalable and continuously growing source of demonstrations for dexterous manipulation learning.

Given one human demonstration video as a query, RoboTok retrieves other web videos with similar underlying manipulation motions.

Rather than matching videos by labels or visual appearance, RoboTok compares how the hands move over time.

💡 The key idea is to represent each video with canonicalized 3D hand trajectories. Each trajectory is expressed in an estimated actor-centered reference frame, so the movement is described relative to the person rather than the camera. This makes similar manipulation behaviors easier to compare across different viewpoints and scenes, even when the actor is partly occluded.

In our experiments, RoboTok retrieved more manipulation-relevant demonstrations than existing robot-data retrieval methods. When those videos were used to guide robot training, the simulated robots completed manipulation tasks more successfully.

I’m sincerely grateful to Howard Qian and Kaiyu Hang for leading this project. I also want to thank Yiting Chen, Kejia Ren, Podshara Chanrungmaneekul, Gaotian Wang, Bowen Wen @bowenwen_me, and Chen Wei @_Chen_Wei_ for their guidance and collaboration.

Project site: https://t.co/v43CCf0040
Paper: https://t.co/WZGwQLhDTK
Code: https://t.co/MfYLTxHopH
Data and models: https://t.co/i0s7RcrXvX

[Johnny Núñez @itsjohnnynunez]

@cindy_x_wu I love it!

[Xindi Wu @cindy_x_wu]

The ICML 2026 oral recordings are now live:)

Check out our talk on motion attribution for video generation, where we study how different training examples influence generated motion and what this can reveal about video generation models.

🔗 https://t.co/5r8JmnQQ0N