Ant Group trained a robot vision-language-action model on 20,000 hours of robot data.
January 29, 2026 at 12:01 AM
Ant Group trained a robot vision-language-action model on 20,000 hours of robot data.
Sharpa has a new vision-tactile-language-action model, and they demoed it live at CES. Video is autonomous and some of the most impressive dexterous manipulation I have ever seen.
January 8, 2026 at 2:50 PM
Sharpa has a new vision-tactile-language-action model, and they demoed it live at CES. Video is autonomous and some of the most impressive dexterous manipulation I have ever seen.
⭐ The first foundational model available on @LeRobotHF ⭐
Pi0 is the most advanced Vision Language Action model. It takes natural language commands as input and directly output autonomous behavior.
It was trained by @physical_int and ported to pytorch by @m_olbap
👇🧵
Pi0 is the most advanced Vision Language Action model. It takes natural language commands as input and directly output autonomous behavior.
It was trained by @physical_int and ported to pytorch by @m_olbap
👇🧵
February 4, 2025 at 5:07 PM
⭐ The first foundational model available on @LeRobotHF ⭐
Pi0 is the most advanced Vision Language Action model. It takes natural language commands as input and directly output autonomous behavior.
It was trained by @physical_int and ported to pytorch by @m_olbap
👇🧵
Pi0 is the most advanced Vision Language Action model. It takes natural language commands as input and directly output autonomous behavior.
It was trained by @physical_int and ported to pytorch by @m_olbap
👇🧵
We don't want vision-language-action models to remain bounded by human performance, but training these large models with reinforcement learning is difficult. How can we handle this? Short blog post: open.substack.com/pub/itcanthi...
Paper notes: Improving Vision-Language-Action Model with Online Reinforcement Learning
Many vision-language-action models really don’t work that well.
open.substack.com
February 6, 2025 at 2:39 PM
We don't want vision-language-action models to remain bounded by human performance, but training these large models with reinforcement learning is difficult. How can we handle this? Short blog post: open.substack.com/pub/itcanthi...
Vision-Language-Action models are the foundation of a new wave of generalist robots: networks that take in images from robot cameras (vision) and instructions (language) and produce robot trajectories. We are seeing a remarkable convergence in how these work; more: open.substack.com/pub/itcanthi...
Vision-Language-Action Models and the Search for a Generalist Robot Policy
VLAs are general-purpose robotics models. But how are VLAs doing in the real world. and which ones are people using?
open.substack.com
August 27, 2025 at 11:42 PM
Vision-Language-Action models are the foundation of a new wave of generalist robots: networks that take in images from robot cameras (vision) and instructions (language) and produce robot trajectories. We are seeing a remarkable convergence in how these work; more: open.substack.com/pub/itcanthi...
📣 Excited to share our #CVPR2025 Spotlight paper and my internship project @wayve: SimLingo.
A Vision-Language-Action (VLA) model that achieves state-of-the-art driving performance with language capabilities.
Code: github.com/RenzKa/simli...
Paper: arxiv.org/abs/2503.09594
A Vision-Language-Action (VLA) model that achieves state-of-the-art driving performance with language capabilities.
Code: github.com/RenzKa/simli...
Paper: arxiv.org/abs/2503.09594
May 8, 2025 at 3:25 PM
📣 Excited to share our #CVPR2025 Spotlight paper and my internship project @wayve: SimLingo.
A Vision-Language-Action (VLA) model that achieves state-of-the-art driving performance with language capabilities.
Code: github.com/RenzKa/simli...
Paper: arxiv.org/abs/2503.09594
A Vision-Language-Action (VLA) model that achieves state-of-the-art driving performance with language capabilities.
Code: github.com/RenzKa/simli...
Paper: arxiv.org/abs/2503.09594
AdaDE enhances vision-language-action policies by converting dense networks into efficient MoE structures, enabling dynamic expert deactivation that retains 95.7% success in complex robotic tasks while reducing model size by 40%. https://arxiv.org/abs/2609.16503
Dense to MoE Adaptation for Compact Vision Language Action Policies
ArXiv link for Dense to MoE Adaptation for Compact Vision Language Action Policies
arxiv.org
September 22, 2026 at 5:20 PM
AdaDE enhances vision-language-action policies by converting dense networks into efficient MoE structures, enabling dynamic expert deactivation that retains 95.7% success in complex robotic tasks while reducing model size by 40%. https://arxiv.org/abs/2609.16503
AdaDE transforms dense vision-language-action policies into compact mixture-of-experts models, achieving 50% parameter reduction while maintaining performance in robotic tasks. It addresses deployment challenges for resource-constrained robotics. https://arxiv.org/abs/2609.16503
Dense to MoE Adaptation for Compact Vision Language Action Policies
ArXiv link for Dense to MoE Adaptation for Compact Vision Language Action Policies
arxiv.org
September 22, 2026 at 2:10 PM
AdaDE transforms dense vision-language-action policies into compact mixture-of-experts models, achieving 50% parameter reduction while maintaining performance in robotic tasks. It addresses deployment challenges for resource-constrained robotics. https://arxiv.org/abs/2609.16503
A bunch of companies are pushing really hard to get on-device inference working for local LLMs, but the thing is, if a cellphone can run a chatbot that same cheap hardware can run vision/language/action models for an automated drone that picks targets based on, oh, idk, regional dialect
My colleague Mr. Sterling reminds me that while I may drop into fond geriatric nostalgia at the drop of a pink plastic carabiner clip, the global street finding its own military uses for user-operated drones has very dire potential.
May 7, 2026 at 6:02 PM
A bunch of companies are pushing really hard to get on-device inference working for local LLMs, but the thing is, if a cellphone can run a chatbot that same cheap hardware can run vision/language/action models for an automated drone that picks targets based on, oh, idk, regional dialect
They find that pretrained Vision-Language-Action (VLA) models are surprisingly resistant to forgetting!
Zero forgetting, or even positive backward transfer, is possible with simple experience replay.
Zero forgetting, or even positive backward transfer, is possible with simple experience replay.
March 10, 2026 at 4:35 AM
They find that pretrained Vision-Language-Action (VLA) models are surprisingly resistant to forgetting!
Zero forgetting, or even positive backward transfer, is possible with simple experience replay.
Zero forgetting, or even positive backward transfer, is possible with simple experience replay.
*That mindless automaton never made a metaphysically-correct moral decision to commit that offense
*To be technically precise, it maneuvered its 6-DoF end-effector in the Vision-Language-Action Model token-space
*To be technically precise, it maneuvered its 6-DoF end-effector in the Vision-Language-Action Model token-space
December 18, 2025 at 7:24 AM
*That mindless automaton never made a metaphysically-correct moral decision to commit that offense
*To be technically precise, it maneuvered its 6-DoF end-effector in the Vision-Language-Action Model token-space
*To be technically precise, it maneuvered its 6-DoF end-effector in the Vision-Language-Action Model token-space
Turning vision language models into vision language action models or leveraging computer vision for robotics
September 16, 2025 at 3:25 PM
Turning vision language models into vision language action models or leveraging computer vision for robotics
This study introduces Direction-Scale Decomposition (DSD) for action representation in vision-language-action models, improving efficiency by separating motion direction from speed. Experiments show DSD boosts robotic manipulation success rates, marking progress. https://arxiv.org/abs/2609.28865
Direction-Scale Decomposition in Action Representation: Rethinking What to Tokenize for Vision-Language-Action Models
ArXiv link for Direction-Scale Decomposition in Action Representation: Rethinking What to Tokenize for Vision-Language-Action Models
arxiv.org
September 25, 2026 at 7:20 PM
This study introduces Direction-Scale Decomposition (DSD) for action representation in vision-language-action models, improving efficiency by separating motion direction from speed. Experiments show DSD boosts robotic manipulation success rates, marking progress. https://arxiv.org/abs/2609.28865
Fast robotics needs fast AI inference.
InstinctFlash is an open-source inference runtime for robotics and Vision-Language-Action models.
Interesting direction for physical AI, real-time robotics and edge inference.
Full breakdown → ToolSpire
#RoboticsAI #PhysicalAI #OpenSource
InstinctFlash is an open-source inference runtime for robotics and Vision-Language-Action models.
Interesting direction for physical AI, real-time robotics and edge inference.
Full breakdown → ToolSpire
#RoboticsAI #PhysicalAI #OpenSource
September 27, 2026 at 4:25 PM
Fast robotics needs fast AI inference.
InstinctFlash is an open-source inference runtime for robotics and Vision-Language-Action models.
Interesting direction for physical AI, real-time robotics and edge inference.
Full breakdown → ToolSpire
#RoboticsAI #PhysicalAI #OpenSource
InstinctFlash is an open-source inference runtime for robotics and Vision-Language-Action models.
Interesting direction for physical AI, real-time robotics and edge inference.
Full breakdown → ToolSpire
#RoboticsAI #PhysicalAI #OpenSource
TANDEM merges task and motion planning with human teleoperation, minimizing human involvement in robot demos. This strategy boosts data collection, yielding a 2.9x increase in successful demos and improving task success rates for effective robot learning. https://arxiv.org/abs/2609.28314
TANDEM: Task and Motion Planning with As-Needed Demonstrations for Efficient Vision-Language-Action Model Fine-tuning
ArXiv link for TANDEM: Task and Motion Planning with As-Needed Demonstrations for Efficient Vision-Language-Action Model Fine-tuning
arxiv.org
September 25, 2026 at 2:20 AM
TANDEM merges task and motion planning with human teleoperation, minimizing human involvement in robot demos. This strategy boosts data collection, yielding a 2.9x increase in successful demos and improving task success rates for effective robot learning. https://arxiv.org/abs/2609.28314
*I'm starting to think that "Vision-Language-Action Models really are "robot imiitators," but they might "imitate" robots better than previous robots have ever been "robots"
July 1, 2025 at 6:26 AM
*I'm starting to think that "Vision-Language-Action Models really are "robot imiitators," but they might "imitate" robots better than previous robots have ever been "robots"
Vision Language Action models are neat. Conceptually they appear to serve as a general solution to what used to be a whole bunch of bespoke robotics problems. I am sure there is much nuance, but seeing the capabilities of coding agents, I'm eager to see embodiment in action. arxiv.org/pdf/2405.14093
August 26, 2025 at 9:29 PM
Vision Language Action models are neat. Conceptually they appear to serve as a general solution to what used to be a whole bunch of bespoke robotics problems. I am sure there is much nuance, but seeing the capabilities of coding agents, I'm eager to see embodiment in action. arxiv.org/pdf/2405.14093
*I'd be mildly alarmed by Vision-Language-Action AI if it wasn't being built by Google, who will promptly strangle any Google initiative that doesn't buy yachts for its shareholders #KilledByGoogle
June 27, 2025 at 6:18 AM
*I'd be mildly alarmed by Vision-Language-Action AI if it wasn't being built by Google, who will promptly strangle any Google initiative that doesn't buy yachts for its shareholders #KilledByGoogle
I have updated my tutorial on making Vision Language Action models. This tutorial starts with a basic Transformer and walks people through the steps to transform it into a full VLA that uses PaliGemma as the pretrained VLM. Links below.
February 9, 2026 at 2:15 PM
I have updated my tutorial on making Vision Language Action models. This tutorial starts with a basic Transformer and walks people through the steps to transform it into a full VLA that uses PaliGemma as the pretrained VLM. Links below.
Can 7B parameters outmaneuver bigger robot brains? FLUX 3 Action bets on speed, openness, and lower costs.
#Robotics #AI #FLUX3 https://spaisee.com/article/black-forest-labs-targets-robotics-with-smaller-faster-flux-3-action
#Robotics #AI #FLUX3 https://spaisee.com/article/black-forest-labs-targets-robotics-with-smaller-faster-flux-3-action
Black Forest Labs Targets Robotics With Smaller, Faster FLUX 3 Action
Black Forest Labs’ FLUX 3 Action is a 7-billion-parameter open-weight robotics model promising faster, cheaper deployment through vision-language control, adaptive feedback, and fine-tuning for real-world machines.
spaisee.com
September 27, 2026 at 1:00 AM
Can 7B parameters outmaneuver bigger robot brains? FLUX 3 Action bets on speed, openness, and lower costs.
#Robotics #AI #FLUX3 https://spaisee.com/article/black-forest-labs-targets-robotics-with-smaller-faster-flux-3-action
#Robotics #AI #FLUX3 https://spaisee.com/article/black-forest-labs-targets-robotics-with-smaller-faster-flux-3-action
Together, we dig into the research behind that idea, from code-as-policies and vision-language-action models to the harnesses and evals needed to make these systems work in the real world.
00:00 — What Are Robot-Use Agents?
02:11 — From Vision-Language-Action Models to General LLMs
00:00 — What Are Robot-Use Agents?
02:11 — From Vision-Language-Action Models to General LLMs
September 26, 2026 at 3:22 PM
Together, we dig into the research behind that idea, from code-as-policies and vision-language-action models to the harnesses and evals needed to make these systems work in the real world.
00:00 — What Are Robot-Use Agents?
02:11 — From Vision-Language-Action Models to General LLMs
00:00 — What Are Robot-Use Agents?
02:11 — From Vision-Language-Action Models to General LLMs
*I suspect that LLMs will become dead-media in pretty short order, replaced by "Large Multimodal Models" and "Vision-Action-Language Models," and then "we" won't be "perpetuating" them any more than "we" "perpetuate" floppy-disks
i'm simply answering that question: how do these perpetuate? We perpetuate them. It isn't good, but people are doing a lot of the heaving lifting of meaning with "AI" at the moment. Their "agency" still needs to propagate and have meaning to stick around.
*Yeah? How are you gonna "deny their agency" when they start murdering each other over precious voltage-shortages 🤖☠️
December 15, 2025 at 7:44 AM
*I suspect that LLMs will become dead-media in pretty short order, replaced by "Large Multimodal Models" and "Vision-Action-Language Models," and then "we" won't be "perpetuating" them any more than "we" "perpetuate" floppy-disks
What can you build with a fully open robotics model in a weekend? 🤖
Robotics engineer @0xbinh.bsky.social used MolmoAct 2, our open vision-language-action model, in the voice-controlled robot that won @southparkcommons.bsky.social's AI hackathon.
Watch our interview with him ↓ 🎥
Robotics engineer @0xbinh.bsky.social used MolmoAct 2, our open vision-language-action model, in the voice-controlled robot that won @southparkcommons.bsky.social's AI hackathon.
Watch our interview with him ↓ 🎥
July 8, 2026 at 6:39 PM
What can you build with a fully open robotics model in a weekend? 🤖
Robotics engineer @0xbinh.bsky.social used MolmoAct 2, our open vision-language-action model, in the voice-controlled robot that won @southparkcommons.bsky.social's AI hackathon.
Watch our interview with him ↓ 🎥
Robotics engineer @0xbinh.bsky.social used MolmoAct 2, our open vision-language-action model, in the voice-controlled robot that won @southparkcommons.bsky.social's AI hackathon.
Watch our interview with him ↓ 🎥
Yeah but language models can't produce AGI
That's why we're adding vision encoders to them and heartlessly throwing them to the RL gulag and maybe soon grafting a language-action decoder to the backbone too
That's why we're adding vision encoders to them and heartlessly throwing them to the RL gulag and maybe soon grafting a language-action decoder to the backbone too
September 15, 2026 at 9:41 AM
Yeah but language models can't produce AGI
That's why we're adding vision encoders to them and heartlessly throwing them to the RL gulag and maybe soon grafting a language-action decoder to the backbone too
That's why we're adding vision encoders to them and heartlessly throwing them to the RL gulag and maybe soon grafting a language-action decoder to the backbone too
Understanding the Impact of Geometric Foundation Models on Vision-Language-Action Models
Yang,Lin,Martin-Martin,Labrie,Gayaka,Kuo,Carlone
Put VGGT into VLA.
Early fusion (before VLM) > late fusion (after VLM)
Impact of 3D (VGGT) highest in single camera setup (!?!)
arxiv.org/abs/2605.24642
Yang,Lin,Martin-Martin,Labrie,Gayaka,Kuo,Carlone
Put VGGT into VLA.
Early fusion (before VLM) > late fusion (after VLM)
Impact of 3D (VGGT) highest in single camera setup (!?!)
arxiv.org/abs/2605.24642
May 27, 2026 at 1:46 PM
Understanding the Impact of Geometric Foundation Models on Vision-Language-Action Models
Yang,Lin,Martin-Martin,Labrie,Gayaka,Kuo,Carlone
Put VGGT into VLA.
Early fusion (before VLM) > late fusion (after VLM)
Impact of 3D (VGGT) highest in single camera setup (!?!)
arxiv.org/abs/2605.24642
Yang,Lin,Martin-Martin,Labrie,Gayaka,Kuo,Carlone
Put VGGT into VLA.
Early fusion (before VLM) > late fusion (after VLM)
Impact of 3D (VGGT) highest in single camera setup (!?!)
arxiv.org/abs/2605.24642