What is a VLA (vision-language-action) model?
A VLA model is a robot foundation model that takes camera images and a language instruction and outputs robot actions directly.
Updated:
Definition
A vision-language-action (VLA) model maps what the robot sees and what it is told to what it should do. Most start from a pretrained vision-language model and add an action head trained on robot data.
How it works
- Encode images (and often robot state) plus the text instruction.
- Predict future actions such as end-effector poses and gripper commands.
- Execute some actions, observe again, and re-plan.
Why it matters
VLAs are the main approach behind "generalist" robot policies that work across tasks and robot types. See embodied AI.
Frequently asked questions
How does a VLA differ from a vision-language model (VLM)?
A VLM answers in text. A VLA adds an action output, typically a short sequence of future motor commands, so it can control a robot.
What is an action chunk?
A fixed-length sequence of future actions predicted at once. Research such as Vela explores continuous trajectories instead.
What are the limitations of VLAs?
Many rely mostly on the current camera frame, lack touch and memory, and struggle with long, contact-rich tasks. Models like NEURA's Neuron AI add more senses and memory.
Coverage
NEURA's Neuron AI adds touch, sound and memory to the robot foundation model
NEURA Robotics launched Neuron AI, a 'perception-language-action' model that fuses 3D vision, touch, force, audio, robot state and language with memory, and will demo an 11-minute laundry task on its 4NE1 humanoid.
Physical Intelligence opens its π0 robot models to everyone via an API
The Pi API beta lets robot builders upload data, fine-tune π0/π0.5 on Pi's infrastructure and run the result through hosted inference.
Unitree's upgraded G1+ humanoid goes on pre-order in Japan
GMO AIR began Japanese pre-orders for the G1+ on Oct 13: a neck, stronger motors, 3D LiDAR and voice interaction. Founder Wang Xingxing now sees an embodied-AI tipping point in 1–2 years.
1X puts its NEO home humanoid into real households for a beta trial
A three-month trial with selected households will test whether the $20,000 NEO can do chores safely and reliably before customer deliveries.
Ultra raises $62M for warehouse robots-as-a-service — with Physical Intelligence as the brain
Brooklyn's Ultra closed a $50M Series A (Framework Ventures) plus earlier seed for $62M total, leasing OP1 packers while Physical Intelligence supplies the AI.
Vela: a VLA that predicts smooth trajectories instead of action chunks
A Fudan/CUHK team's Vela model outputs spline curves with an adaptive time horizon, beating π0.5 on LIBERO-X and EBench and on real egg-cake cooking and potato-shredding tasks.
Skild's S1 learns soccer by playing itself for 140 simulated years
After launching S1 — a robotics foundation model that copies a single video demo without fine-tuning — Skild showed Physical Self-Play: the same base policy trained against itself in Isaac Sim, then transferred to a humanoid.
Figure's Helix 2.5 cleans 30 homes it has never seen
Figure says its humanoid tidied rooms, folded towels and made beds in 30 unseen Bay Area homes with zero local data, backed by a $3.5B compute deal with Nscale.
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.