TryThink retweeted
Happy to share EgoWAM, led by @BaoyuLi6!
Naive co-training on in-the-wild human data hurts policy performance. The reason is simple: humans move very differently from robots.
EgoWAM bridges this gap through world representations: humans and robots may move differently, but achieve similar effects. We find that motion-centric representations and strong visual pretraining solve different parts of the embodiment gap.
Egocentric human data is abundant, but human motion is not always positive supervision for robot policy due to embodiment gaps. Naive BC co-training can HURT performance ☹️.
🌟Our key finding in **EgoWAM**: the state-prediction branch of a World Action Model effectively bridges this embodiment gap, enabling robot performance to scale with diverse **in-the-wild** human data.
💡The key question then becomes: what world representation transfers best across embodiments?
👇🏻Let’s take a deep dive into it:
🌐 gatech-rl2.github.io/egowam.…
🧵[1/]
TryThink retweeted
*NEW* Our #ECCV2026 @eccvconf paper
Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation
Now on ArXiv w Dataset, Code&model
sid2697.github.io/epic-conta…
arxiv.org/abs/2606.30598
Two contributions:
1. EPIC-Contact Dataset
2. HOPformer Method &Checkpoint
🧵 1/6
TryThink retweeted
Unlike symbolic systems, robotics doesn't have the luxury of a "common abstraction" between human and machines.
Robotics instead speak in the language of physics: contact, force, wrench.
Happy to share another step in human2robot transfer for dexterous manipulation, using contact wrench as a bridge.
Led by @zhu_xinghao and @Dr_YanChang 's team at NVIDIA.
How can robots learn dexterous manipulation from human demonstrations at scale?
Excited to share CHORD: Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration.
CHORD learns from human demos by focusing not only on where contact happens, but how that contact moves the object through force and torque guidance.
This unified contact-wrench representation carries human manipulation skills across diverse behaviors, long-horizon tasks, whole-body embodiments, and real-world hardware.
We evaluated CHORD on large-scale, long-horizon, contact-rich tasks paired with human demonstrations, spanning rigid, articulated, and multi-object manipulation.
At scale:
* 82.12% average success across 1,831 tasks
* 90.77% whole-body manipulation success
* 4,739 sim-ready dexterous manipulation benchmark
* Transfer to real dexterous hands
Project page: nvidia-isaac.github.io/video…
Tech report: nvidia-isaac.github.io/video…
Code will be released soon as part of Video to Data repo github.com/nvidia-isaac/vide…, our end-to-end pipeline for converting human demonstration videos into simulation-ready assets and physics-grounded robot training data.
Huge thanks to amazing contributors: @zhu_xinghao , Zixi Liu, Shalin Jain, Chenran Li, Milad Noori, Huihua Zhao, John Welsh, @michaelv03, Wei Liu, @TingwuWang , Xingye (Dennis) Da, @zhengyiluo, Vishal Kulkarni, @sNaema, @yukez, @DrJimFan, @bowenwen_me, @danfei_xu, @SohaPouya, @Dr_YanChang.
#Robotics #PhysicalAI #DexterousManipulation #RobotLearning #NVIDIA
TryThink retweeted
Introduce SARM2 🤖a multi-task stage-aware reward model that empowers a self-improving loop:
🧺 Folding Shorts 58% → 100%
🧽 Cleaning Whiteboard 50% → 90%
Paper + project page below 👇 (1/n)
TryThink retweeted
The robotics community has collected enormous robot datasets on grippers, but what happens when the hardware changes? How do we make robot data outlive the hardware it was collected on?
We introduce Cloak, a training recipe for zero-shot cross-embodiment transfer. We never collect ANY embodiment-specific data. 🧵
TryThink retweeted
Introduce EgoInfinity:
a web-scale (14.6 yrs, 142M clips)
data engine that automatically lifts Youtube videos into 4D hand-object interaction, and retargets for robot learning
Not only a dataset, but a modular, upgradable engine
HF Space: huggingface.co/spaces/Rice-R…
TryThink retweeted
World-Action Models (WAMs) have become the second dominant recipe for robot foundation models, next to classical VLAs.
So where do they come from, and how do they compare vs VLAs?
I wrote an small overview of the WAM landscape, with some personal takes:
developer.nvidia.com/blog/pr…
TryThink retweeted
Hot take: robots should not dream in pixels.
Pixels are too low-level.
Latents are too opaque.
μ₀ predicts a third thing:
3D motion traces.
On real robots, it beats π₀.₅ — with ~1/100 the data scale and no action labels for world-model pretraining. 🧵
mu0-wm.github.io/
(this video features voiceover narration)
TryThink retweeted
π, But Make It Fly ✈️
We fine-tuned π0, a VLA model pretrained entirely on manipulators, to fly a drone that picks up objects, navigates through gates, and composes both skills from language commands.
TryThink retweeted
1X NEO is an adult-human-sized humanoid that can be lifted relatively easily. Being lightweight is a key safety feature.
TryThink retweeted
🚀 Low-cost whole-body teleoperation device released for Vega robot!
🔓 Open-source
💰 Low-cost
🎒 Portable
🦾 Whole-body control
See it live at our #CoRL booth! 🤖✨
🚀 Launching our open-source teleoperation device for full-body Vega robot control!
🔓 Open-source
💰 Low-cost
🎒 Portable
🦾 Whole-body control
Control both real and simulated Vega robots! 🤖✨
📍 Live Demo at #CoRL - See it in action with real and simulated Vega robots in Isaac Lab. Stop by booth #6 if you're attending!
#PhysicalAI #Robotics #Teleoperation
TryThink retweeted
Can robotic hands really outperform human dexterity? Meet the Linker Hand L30 from LinkerBot. This tendon-driven dexterous hand packs high precision and compact size, making it a serious contender for complex manipulation tasks:
► 22 Degrees of Freedom utilizing a sophisticated tendon drive system.
► Ultra-fine positioning control with ±0.20mm repeatable accuracy.
► Rapid response speed, achieving up to 450°/S for core movements.
► Adaptive grasping enabled by high-sensitivity tactile sensors built into the fingertips.
► Compact dimensions (85mm average width, 179mm length) despite its complex architecture.
Linkerbot Unveils a Lightweight, High-Performance Dexterous Hand Priced Under $1K
Linkerbot has just launched the Linker Hand O6, a new cost-effective and high-performance dexterous hand. It's designed to be lightweight and compact, weighing just 370g—around the size of a female hand.
► The hand features a total of 11 degrees of freedom (6 active, 5 passive).
► It utilizes new aerospace-grade materials, giving it an extremely low weight and an impressive payload capacity of over 30kg.
► Priced at just 6,666 RMB (approximately $912 USD), the Linker Hand O6 is positioned as a "small yet powerful" high-performance, cost-effective option.
TryThink retweeted
DexWrist got an upgrade! Robots with wrists can do so much more! dexwrist.csail.mit.edu/
Kudos to @martinpeticco @JohnMarangola
TryThink retweeted
Large-scale 3D Scene Generation (all scenes are real-time rendered)!!
Physically-grounded generative data without hallucinations is the missing link for robot learning and testing at scale. We introduce a method that directly generates large-scale 3D driving scenes with accurate geometry, allowing for causal view synthesis and generation with object permanence and explicit 3D geometry. This also allows for extreme trajectory extrapolation without failure! We also show that we can build fully data-driven simulators for end-to-end learning with this approach.
Project: light.princeton.edu/lsd-3d/
with the amazing team of Julian Ost, @amogh7joshi , Andrea Ramazzina, Maximilian Bömer, Mario Bijelic.
TryThink retweeted
Anthropic’s CEO just gave the most important AI interview of 2025.
He revealed shocking truths about artificial super intelligence, the OpenAI rivalry, and why AI is advancing at a pace no one saw coming.
He even called out Zuckerberg.
Here are my top 8 takeaways: 🧵
TryThink retweeted
Intelligent Robot Simulator (IR-SIM)
A Python based lightweight robot simulator for the development of algorithms in robotics navigation, control, and learning.
github.com/hanruihua/ir-sim
