Train
Generate reward signals from robot experience so learning can scale beyond hand-written functions and sparse human feedback.
About Sublinear
Sublinear is building foundation models that turn robot video, natural-language goals, and safety criteria into feedback about progress, success, failure, and risk.
Robot experience
Video + task goal
Foundation model
Progress
Success
Failure
Safety
The feedback gap
More robot data does not automatically create better robots.
Robotics teams collect enormous volumes of video, but the robot cannot inherently tell whether an action succeeded, failed, moved toward a goal, or created an unsafe situation. Without that understanding, much of the experience remains difficult to learn from.
Today, those judgments usually come from human labels, which are slow and expensive.
Every trajectory raises questions
- Did the robot complete the task?
- How much progress did it make?
- Where did the behavior break down?
- Did anything unsafe happen?
A general capability
Evaluating behavior should be intelligence, not repeated manual labor.
Our models provide a shared way to interpret robot experience. Given visual observations and a description of the goal, Sublinear produces dense progress rewards and success estimates. Safety criteria can be expressed in language too, making it possible to identify behavior that deserves attention.
The result is one underlying intelligence layer that can support learning and safety evaluation across the robot lifecycle.
Evaluate
Measure progress and success across policies, tasks, environments, and robot embodiments with a consistent interface.
Monitor safety
Apply natural-language safety criteria to surface risky behavior, failures, and episodes that need human review.
Improve data
Score, rank, filter, and clean large video datasets so the most useful experiences become easier to find and reuse.
Built from research
A research problem becoming a systems problem.
Sublinear is built upon over three years of PhD research at MIT spanning reinforcement learning, video-language reasoning, and foundation reward models. That work demonstrated the core approach on real robot behavior.
We are now turning that research into practical infrastructure for robotics companies, academic and industry labs, and developers building learning, evaluation, and safety systems.
Help robots learn from every experience.
Use Sublinear to turn robot video into signals for training, evaluation, data curation, and safety review.