It turns out you can just take an off-the-shelf VLM and fine-tune it directly to output robot actions *as text* and it performs better than/as-good-as all the more complex model architectures…
Check out the paper!
What's the right architecture for a VLA?
VLM + custom action heads (π₀)?
VLM with special discrete action tokens (OpenVLA)?
Custom design on top of the VLM (OpenVLA-OFT)?
Or... VLM with ZERO modifications? Just predict action as text.
The results will surprise you.
VLA-0: Outperforms π₀, GR00T-N1, MolmoAct, SmolVLA.
With ZERO changes to the VLM.
🧵⬇️
There is a made-up letter circulating online asking NVIDIA’s board to reinstate Jensen, claiming my involvement. This letter is obviously fake, and even if it were real I would not sign it.
I am grief-stricken that Jensen is no longer with NVIDIA. At the same time, I fully support NVIDIA’s board and have confidence in their decisions.
I’ve worked with Colette for years and know her as an extraordinary leader. She has my full support.
Today we are introducing ARC, a reasoning recipe for robot foundation models that sets a new state of the art on RoboLab-120 and MolmoSpaces.
The prevailing way to improve robot foundation models is brute force: larger models, more robot demonstrations, costly training at scale. ARC needs none of that. With no new robot data and no foundation-scale training, π0.5 and Cosmos3-Nano-Policy gain up to:
📈 +50.0 pp on RoboLab-Reasoning-50
📈 +29.8 pp on RoboLab-120
📈 +27.2 pp on MolmoSpaces
To our knowledge, these gains are unprecedented.
📄 Paper: arxiv.org/abs/2610.12386
🌐 Website: arc-robot-reasoning.github.i…
🧵 (1/7)
Excited to share that VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation has been accepted to #CoRL2026! 🎉🤖
VoLo explores agentic robotics for long-horizon, open-vocabulary tasks, where a VLM acts as the robot’s brain with memory to reason, plan, monitor, detect failures, recover, and orchestrate tools—including VLA/WAMs, vision models, and action primitives.
We also introduce RoboVoLo, a high-fidelity simulation benchmark with 126 tasks spanning common sense, memory/state tracking, complex references, and world knowledge, together with detailed diagnosis of current system limitations.
This work was done during my internship at @nvidia@NVIDIAAI@NVIDIARobotics. Huge thanks to all collaborators @HugoHadfield1, Alexander Zook, @mikacuy, @luke_ch_song, @erwincoumans, @xuningy, Faisal Ladhak, @BirchfieldStan, Jonathan Tremblay, and @robovalts!
🔗 Project: chicychen.github.io/VoLo/
💻 Code: github.com/NVlabs/VoLoAgent
📊 RoboVoLo Benchmark: github.com/NVlabs/RoboVoLo
Looking forward to discussions at CoRL 2026 — hope to meet many of you in person!
#CoRL2026#AgenticRobotics#Robotics#EmbodiedAI#RobotLearning#PhysicalAI#NVIDIA
We have just open sourced Droid+, a really simple and easy way to drive the Droid hardware system using modern web technologies and very few dependencies! I hope people find it useful! github.com/NVlabs/DroidPlus
We released DROID+, a new and improved DROID control stack from SRL. We believe robots (and hardware) should be easy to setup and easy to use, and so we updated the DROID stack to be the same. @HugoHadfield1github.com/nvlabs/droidplus
Here's pi0.5 doing our "hello world" task:
We're hiring! The Seattle Robotics Lab (SRL) at NVIDIA is hiring a research scientist to work on robot foundation models. The focus will be on investigating the science behind building VLAs, WAMs, and world models for robotics, and leveraging the findings to propose the next generation of models and training paradigms.
We're seeking a world-class researcher (PhD grad) or senior researcher (3+ years of post-PhD experience) with expertise in training generalist policies. The researcher should aim to question conventional wisdom, push beyond the status quo, and dive into the foundations of general-purpose robot intelligence.
NVIDIA SRL provides a rare environment to conduct open research on fundamental and applied robotics, with the freedom to publish, collaborate broadly, and tackle big scientific questions at scale.
If this sounds like a strong fit, please reach out to us and apply here!
nvidia.eightfold.ai/careers/…nvidia.eightfold.ai/careers/…
We're hiring! The Seattle Robotics Lab (SRL) at NVIDIA is hiring a research scientist to work on robot foundation models. The focus will be on investigating the science behind building VLAs, WAMs, and world models for robotics, and leveraging the findings to propose the next generation of models and training paradigms.
We're seeking a world-class researcher (PhD grad) or senior researcher (3+ years of post-PhD experience) with expertise in training generalist policies. The researcher should aim to question conventional wisdom, push beyond the status quo, and dive into the foundations of general-purpose robot intelligence.
NVIDIA SRL provides a rare environment to conduct open research on fundamental and applied robotics, with the freedom to publish, collaborate broadly, and tackle big scientific questions at scale.
If this sounds like a strong fit, please reach out to us and apply here!
nvidia.eightfold.ai/careers/…nvidia.eightfold.ai/careers/…
Wonderful to be back from #CVPR2026, and excited to share the release of our follow-up work:
VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation
VoLo introduces the idea of a physical orchestrator for open-vocabulary, long-horizon manipulation. Our goal is to move toward robots that can reason, plan, act, monitor, and recover by adaptively using VLA/WAMs, vision models, and action primitives as tools.
We introduce three main contributions:
🤖 VoLoAgent — a physical orchestrator that plans, monitors, and recovers by adaptively using, halting, and redirecting robot actions with tools.
📊 RoboVoLo — a high-fidelity benchmark with 126 open-vocabulary long-horizon manipulation tasks spanning common sense, memory/state tracking, complex references, and world knowledge.
📈 A large-scale empirical study comparing action models, code-as-policy systems, TAMP-style systems, and ablations of the VoLoAgent orchestrator, complemented by real-robot experiments.
This work was done during my internship at @NVIDIA and would not have been possible without my brilliant collaborators: Hugo Hadfield, Alexander Zook, @mikacuy, @luke_ch_song, @erwincoumans, @xuningy, Faisal Ladhak, @qu_1006, @BirchfieldStan, Jonathan Tremblay, and @robovalts. Huge thanks to everyone!
🔗 Project: chicychen.github.io/VoLo/
🔗 Previous work, SpaceTools: spacetools.github.io/#Robotics#EmbodiedAI#VisionLanguageModels#VLAModels#RobotLearning#NVIDIA#CVPR2026#LongHorizonManipulation#AI#ComputerVision
Evaluation is a critical bottleneck in building robot foundation models. Check out our latest work RoboLab, led by @xuningy, which addresses this exact challenge.
Its a high-fidelity simulation environment for testing these models. A truly generalist policy should be able to complete these tasks zero-shot, and this benchmark highlights exactly how far we still have to go. More info 👇
When every generalist robot model scores 95%+ on a benchmark, the numbers become meaningless.
What if we built a photorealistic benchmark that never saturates and can generate new scenes and tasks with AI Workflows in minutes?
We introduce RoboLab! 🧵(1/6)
When every generalist robot model scores 95%+ on a benchmark, the numbers become meaningless.
What if we built a photorealistic benchmark that never saturates and can generate new scenes and tasks with AI Workflows in minutes?
We introduce RoboLab! 🧵(1/6)
What's the right architecture for a VLA?
VLM + custom action heads (π₀)?
VLM with special discrete action tokens (OpenVLA)?
Custom design on top of the VLM (OpenVLA-OFT)?
Or... VLM with ZERO modifications? Just predict action as text.
The results will surprise you.
VLA-0: Outperforms π₀, GR00T-N1, MolmoAct, SmolVLA.
With ZERO changes to the VLM.
🧵⬇️