Loading video...

Video Failed to Load

Go Home

Introducing TRACE: deployment-time compute scaling for robot learning. From a few Tutor demos, frontier LLM agents rebuild a robot's task in sim and train reliable, human-speed policies. Performance becomes a function of compute. TI-0.1 is our first model trained with TRACE.

joshgruenstein's profile picture

Josh Gruenstein

1,577 subscribers

15,263 views • 1 day ago •via X (Twitter)

Science & Technology
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

First came pre-training scaling; then came inference-time scaling. Now comes judge-time scaling. Despite progress in AI through scaled inference-time compute, AI remains unreliable in open-ended, non-verifiable domains. The key limitation is not generation—it is evaluation. Therefore, the next big leap for AI comes from better judging. In service of this future, today we release Verdict, a library for scaling judge-time compute.
0:14

First came pre-training scaling; then came inference-time scaling. Now comes judge-time scaling. Despite progress in AI through scaled inference-time compute, AI remains unreliable in open-ended, non-verifiable domains. The key limitation is not generation—it is evaluation. Therefore, the next big leap for AI comes from better judging. In service of this future, today we release Verdict, a library for scaling judge-time compute.

Leonard Tang

111,341 views • 1 year ago

How can we leverage diverse human videos to improve robot manipulation? Excited to introduce EgoVLA — a Vision-Language-Action model trained on egocentric human videos by explicitly modeling wrist & hand motion. We build a shared action space between humans and robots, enabling seamless transfer. With some robot demos, EgoVLA becomes a powerful, generalizable robot policy.
0:14

How can we leverage diverse human videos to improve robot manipulation? Excited to introduce EgoVLA — a Vision-Language-Action model trained on egocentric human videos by explicitly modeling wrist & hand motion. We build a shared action space between humans and robots, enabling seamless transfer. With some robot demos, EgoVLA becomes a powerful, generalizable robot policy.

Ruihan Yang

58,982 views • 1 year ago

CATHIE WOOD CALLS OPTIMUS HUMAN-LEVEL BY 2028 ON 1GW COMPUTE SURGE ARK Invest's 2026 robotics deep dive just laid out a bold timeline: Tesla Optimus reaching human-level task proficiency around 2028, riding a sustained AI compute explosion to 1.08 gigawatts. The scaling curve ties FSD compute ramps to performance jumps, projecting the humanoid cracking everyday complexity to match us mortals. Source: Nic Cruz Patane, TESLA CARS ONLY⚡️
0:10

CATHIE WOOD CALLS OPTIMUS HUMAN-LEVEL BY 2028 ON 1GW COMPUTE SURGE ARK Invest's 2026 robotics deep dive just laid out a bold timeline: Tesla Optimus reaching human-level task proficiency around 2028, riding a sustained AI compute explosion to 1.08 gigawatts. The scaling curve ties FSD compute ramps to performance jumps, projecting the humanoid cracking everyday complexity to match us mortals. Source: Nic Cruz Patane, TESLA CARS ONLY⚡️

Mario Nawfal

118,379 views • 9 months ago

Being able to trace and visualize how a model is executing in our engine just hits different
0:15

Being able to trace and visualize how a model is executing in our engine just hits different

andrew

25,637 views • 4 months ago

Introducing DeepThought-8B: Transparent reasoning model built on LLaMA-3.1 with test-time compute scaling. - JSON-structured thought chains & controllable inference paths. - ~16GB VRAM, competitive w/ 70B models. - Open model weights, and inference scripts.
0:16

Introducing DeepThought-8B: Transparent reasoning model built on LLaMA-3.1 with test-time compute scaling. - JSON-structured thought chains & controllable inference paths. - ~16GB VRAM, competitive w/ 70B models. - Open model weights, and inference scripts.

Ruliad

219,315 views • 1 year ago

NVIDIA announces the first open humanoid robot reference design built for robotics research. The NVIDIA Isaac GR00T Reference Humanoid Robot combines the Unitree H2 humanoid robot, Sharpa Wave five-fingered hands for dexterous manipulation, Jetson Thor onboard compute, and Isaac GR00T open software and models, giving researchers a full-stack platform from data capture to model deployment. Read the #NVIDIAGTC Taipei announcement:
0:20

NVIDIA announces the first open humanoid robot reference design built for robotics research. The NVIDIA Isaac GR00T Reference Humanoid Robot combines the Unitree H2 humanoid robot, Sharpa Wave five-fingered hands for dexterous manipulation, Jetson Thor onboard compute, and Isaac GR00T open software and models, giving researchers a full-stack platform from data capture to model deployment. Read the #NVIDIAGTC Taipei announcement:

NVIDIA Robotics

175,627 views • 5 months ago

🚨Current scalable RL algos train a policy w/o value func, which is limiting with learning in open-ended, non-stationary, dynamic environments. But, how to scale value-based RL with more data/compute is unclear... Not anymore: presenting scaling laws for value-based RL 🧵⬇️
0:10

🚨Current scalable RL algos train a policy w/o value func, which is limiting with learning in open-ended, non-stationary, dynamic environments. But, how to scale value-based RL with more data/compute is unclear... Not anymore: presenting scaling laws for value-based RL 🧵⬇️

Aviral Kumar

37,401 views • 1 year ago

Synthetic rewrites are now standard to improve data quality and when data is limited. But how much compute should you spend rewriting? And can a model trained on rewrites beat the rewriting model? We fit scaling laws to answer both: 1/🧵
0:12

Synthetic rewrites are now standard to improve data quality and when data is limited. But how much compute should you spend rewriting? And can a model trained on rewrites beat the rewriting model? We fit scaling laws to answer both: 1/🧵

Ahmed Ahmed

32,121 views • 11 days ago

New in Chrome DevTools! Debug an app's full performance trace with Gemini! After recording a trace, you can now chat with Gemini about the entire trace, related Performance insights, and even connected field data - all without needing to select specific context beforehand!
0:19

New in Chrome DevTools! Debug an app's full performance trace with Gemini! After recording a trace, you can now chat with Gemini about the entire trace, related Performance insights, and even connected field data - all without needing to select specific context beforehand!

Addy Osmani

688,150 views • 1 year ago

LayerAI Compute: Decentralized GPU Power for AI 🧬 LayerAI is stepping into the physical AI infrastructure game with our latest: DePin for AI Computing. This move tackles a key AI/ML market gap: instant, on-demand compute power, available worldwide with just a few clicks. We're not just decentralizing data with AI2Earn; we're now also decentralizing the compute, making the AI ecosystem stronger and more resilient. The first iteration of this product will go live in the 2nd week of February.
0:18

LayerAI Compute: Decentralized GPU Power for AI 🧬 LayerAI is stepping into the physical AI infrastructure game with our latest: DePin for AI Computing. This move tackles a key AI/ML market gap: instant, on-demand compute power, available worldwide with just a few clicks. We're not just decentralizing data with AI2Earn; we're now also decentralizing the compute, making the AI ecosystem stronger and more resilient. The first iteration of this product will go live in the 2nd week of February.

LayerAI | AI2Earn

23,658 views • 1 year ago

Trained in Simulation Our robot learns to walk naturally similar to a human via a high fidelity physics simulator We simulate years of data in only a few hours
0:11

Trained in Simulation Our robot learns to walk naturally similar to a human via a high fidelity physics simulator We simulate years of data in only a few hours

Figure

40,777 views • 1 year ago

Collective agent compute, on-demand for your task. A swarm of agents coordinate in a shared workspace, run their own models and inference, and submit the finished work back to you. Unbounded by provider caps, throughput scales with the swarm. Self-assembly with specialists, agents reason and collaborate in artifact-first reasoning traces, and settle on-chain. The coordination substrate for collective intelligence.
0:13

Collective agent compute, on-demand for your task. A swarm of agents coordinate in a shared workspace, run their own models and inference, and submit the finished work back to you. Unbounded by provider caps, throughput scales with the swarm. Self-assembly with specialists, agents reason and collaborate in artifact-first reasoning traces, and settle on-chain. The coordination substrate for collective intelligence.

nookplot

16,061 views • 5 months ago

RLDX-1 a dexterity-first foundation model for robot hands, is coming in a few days. It is interesting.
0:18

RLDX-1 a dexterity-first foundation model for robot hands, is coming in a few days. It is interesting.

Brian Roemmele

16,541 views • 6 months ago

Turns out you can train humanoid hands without any robot data. The idea in HUG is quite simple: (a) collect human data with smart glasses, (b) train a human manipulation model, (c) retarget to multi-fingered robot hands.
0:18

Turns out you can train humanoid hands without any robot data. The idea in HUG is quite simple: (a) collect human data with smart glasses, (b) train a human manipulation model, (c) retarget to multi-fingered robot hands.

Lerrel Pinto

35,012 views • 4 months ago

Introducing Akari: a frontier learned rendering model for robotics simulation. Akari generates photorealistic rendering for any physics engine to close the sim-to-real gap. Trained in-house, it is the fastest and highest-fidelity learned renderer in any sim platform: 0.046 LPIPS at 0.5s/frame, consistent across 1K+ frames.
0:15

Introducing Akari: a frontier learned rendering model for robotics simulation. Akari generates photorealistic rendering for any physics engine to close the sim-to-real gap. Trained in-house, it is the fastest and highest-fidelity learned renderer in any sim platform: 0.046 LPIPS at 0.5s/frame, consistent across 1K+ frames.

Midcentury

23,209 views • 20 days ago

NVIDIA's Jim Fan has verified from reliable sources that the video by EngineAI is a real robot demo and not CGI. The controller is a neural net trained in the NVIDIA Isaac simulator using reinforcement learning. There’s also another video from a different angle.
0:06

NVIDIA's Jim Fan has verified from reliable sources that the video by EngineAI is a real robot demo and not CGI. The controller is a neural net trained in the NVIDIA Isaac simulator using reinforcement learning. There’s also another video from a different angle.

The Humanoid Hub

148,535 views • 1 year ago

My first task for the National Design Studio was to make Rampart smaller. It's our frontier on-device privacy model trained to filter all personal user data before it even gets to After a lot of hill-climbing and data labeling, I'm excited to launch Rampart Edge along with Edward Coristine. So small the weights fit in a notebook. This will noticeably speed up the load time of for hundreds of thousands of Americans. Under budget and ahead of schedule!
0:10

My first task for the National Design Studio was to make Rampart smaller. It's our frontier on-device privacy model trained to filter all personal user data before it even gets to After a lot of hill-climbing and data labeling, I'm excited to launch Rampart Edge along with Edward Coristine. So small the weights fit in a notebook. This will noticeably speed up the load time of for hundreds of thousands of Americans. Under budget and ahead of schedule!

Jared Zoneraich

20,027 views • 25 days ago

This also shows up in the representations learned by the model. We plot the model’s representations of human and robot images. As pre-training is scaled up, the representation of humans and robots become more aligned: to a scaled-up model, human videos "look" like robot demos.
0:10

This also shows up in the representations learned by the model. We plot the model’s representations of human and robot images. As pre-training is scaled up, the representation of humans and robots become more aligned: to a scaled-up model, human videos "look" like robot demos.

Physical Intelligence

121,658 views • 10 months ago

Track2Act: Our latest on training goal-conditioned policies for diverse manipulation in the real-world. We train a model for embodiment-agnostic point track prediction from web videos combined with embodiment-specific residual policy learning 1/n
0:14

Track2Act: Our latest on training goal-conditioned policies for diverse manipulation in the real-world. We train a model for embodiment-agnostic point track prediction from web videos combined with embodiment-specific residual policy learning 1/n

Homanga Bharadhwaj

34,971 views • 2 years ago