Voxel51 has launched model 3.0 of its information platform, including new instruments designed to assist robotics and bodily AI builders handle the large portions of digicam, video and sensor information generated by machines working in the actual world.
The essential downside Voxel51 is making an attempt to unravel is simple: robots want massive quantities of real-world information to coach and enhance the AI fashions that permit them to understand and work together with their environment.
A robotic would possibly gather video from a number of cameras alongside lidar, location data and different sensor readings. Builders then want to look via that data, determine helpful examples, label what occurred, and determine which information ought to be used to coach or check their AI fashions.
As fleets develop, doing this manually turns into more and more tough.
Voxel51 3.0 brings a number of of these duties collectively in a single platform, enabling builders to look, set up, label and consider information collected from robots and different bodily AI methods.
Brian Moore, CEO of Voxel51, says: “We noticed this play out with our prospects constructing autonomous methods and notion fashions. The groups that obtained their information infrastructure proper from the beginning had been capable of transfer quicker.
“Bodily AI and robotics is at that very same inflection level. The businesses constructing that basis at present are going to have an actual benefit as their information and fashions scale.”
One of many important additions is Voxel51 Agent, an AI assistant designed to automate a few of the work concerned in managing robotics datasets.
Quite than merely answering questions via a chat interface, the agent can carry out duties throughout the event course of, together with discovering related information, deciding on helpful examples, labeling information and evaluating outcomes.
For instance, a robotics developer coping with lots of of hours of recorded sensor information would possibly want to seek out events when a robotic encountered a selected object or uncommon state of affairs. Voxel51’s search and AI instruments are designed to make these examples simpler to find with out engineers manually reviewing complete recordings.
Adonai Vera, product supervisor at Voxel51, says: “Bodily AI groups don’t want one other chat interface sitting on high of their information. They want an agent that may keep context throughout the mannequin growth lifecycle and orchestrate actions throughout the complete information flywheel.”
Discovering the vital moments
Voxel51 3.0 additionally introduces an expanded search system combining semantic search, typical textual content queries and metadata filters.
Its multimodal time-series capabilities permit builders to view synchronized streams from completely different sensors and seek for explicit moments inside lengthy recordings.
That is notably related to robotics as a result of a single machine can concurrently generate data from cameras, lidar and different sensors. Understanding an occasion could require engineers to look at a number of of these sources collectively.
The platform additionally contains annotation instruments for 2D pictures, 3D information and video. Its Agentic Labeling functionality can use vision-language AI fashions to automate a few of the technique of figuring out and describing objects and occasions contained within the information.
Voxel51 has additionally added instruments for monitoring the place information comes from and the way it adjustments. Exercise logs, enhancing histories and dataset versioning permit builders to determine which data was used to coach a selected mannequin.
The corporate says its structure is model-agnostic, permitting prospects to make use of their very own basis fashions or specialised AI fashions for duties together with inference, automated labeling and analysis.
Turning robotic expertise into coaching information
The broader thought behind the platform is that have gathered by robots in the actual world might be fed again into the event course of.
If a robotic encounters an uncommon object, lighting situation or state of affairs that causes its AI system to battle, builders can determine that occasion within the recorded information, label and analyze it, and doubtlessly add it to the coaching information used to enhance subsequent variations of the mannequin.
As extra robots function for longer durations, nonetheless, the quantity of recorded information grows quickly. Voxel51 argues that the flexibility to determine the small proportion of helpful coaching examples inside these massive datasets is changing into an vital a part of bodily AI growth.
In response to the corporate’s State of Bodily AI 2026 report, groups it classifies as distinctive are thrice extra probably than struggling groups to spend greater than half of their undertaking time working with information.
Voxel51 3.0 is meant to help that course of from the preliminary uncooked recordings via to searchable and labeled datasets that can be utilized for AI coaching and analysis.
Voxel51 companions with Nebius
Voxel51 can be working with AI infrastructure firm Nebius to mix its information administration expertise with cloud computing assets.
The businesses are internet hosting hands-on workshops the place builders can run Voxel51 workflows utilizing Nebius AI Cloud.
The partnership addresses two associated necessities of bodily AI growth: managing the big portions of information generated by machines and offering the computing energy required to course of that data and practice AI fashions.
Evan Helda, head of bodily AI at Nebius, says: “Bodily AI is the following main frontier for AI as a result of it brings digital intelligence into the world of atoms.
“However getting there’ll take greater than higher fashions. Robotics groups want the compute, cloud platform, and information infrastructure to show each deployment, failure and edge case right into a quicker subsequent iteration. That’s the sort of basis we’re constructing with companions like Voxel51.”
Voxel51 says its platform is utilized by corporations together with Berkshire Gray, Ford, Porsche and Precision Planting.
The brand new Voxel51 3.0 platform can be obtainable via a free self-service model for builders intending to start working with the system.
