Natural speed
Data is captured from people performing tasks at natural working speed, preserving tempo, motion patterns and transitions.
Technology
One model, trained on human data, running robots at human speed.
Mind-1 is MindOn's Physical AI model: an omni-brain that unifies task understanding and robot control. The same model runs fixed dual-arm robots, wheeled dual-arm robots and humanoids.
- 3
- robot embodimentsDual-arm, wheeled dual-arm, humanoid
- 32 ms
- model inferenceDown from 82 ms on Mind-0
- 77.5 ms
- camera to actionDown from 153 ms, end to end
- 1 mm
- data accuracyHuman motion capture
Human efficiency
Task cycle time, against a person.
Many embodied AI systems can complete a task, but at a fraction of human speed. We measure the time to complete the whole task, start to finish. Shorter is better.
Packaging
Parcel sorting
Folding
End-to-end task cycle time per item. Bars are scaled within each task. Source: Mind-1 announcement.
One model, multiple robots
Intelligence, decoupled from the body.
Different robots trade off mobility, dexterity, precision, payload and cost. Real deployments will mix them, so the intelligence should not have to be rebuilt for each one.
High-level intelligence
Understand, reason, decide
A multimodal model performs scene understanding, task reasoning and behaviour generation, and produces temporally consistent trajectory intent.
Low-level control
Execute on the robot
A whole-body controller turns intent into fast, smooth motion and handles each robot's own dynamics, balance and tracking.
Fixed dual-armWheeled dual-armHumanoid
Human-centric data
Learning from how people do the task.
Robot-collected data is expensive, tied to one robot, and slow: operators work around the robot's limits, so the model learns cautious motion. Human data captures the task itself, at natural working speed, and is available at far larger scale.
1 mm accuracy
The in-house collection system captures human manipulation at scale with 1 mm data accuracy.
Automated filtering
A data quality pipeline selects demonstrations that are both efficient and suitable for robot learning.
From human behaviour to robot execution
Four components bridge the gap.
A human body and a robot differ in kinematics, dynamics and workspace, so direct imitation is not enough.
Cross-embodiment data pipeline
Converts large-scale human demonstrations into robot-executable action representations across different robot bodies.
Whole-body action foundation model
Learned from tens of thousands of hours of motion capture, it gives one low-level interface for tracking motion while keeping balance.
Sub-3 cm end-effector tracking accuracy
Real-world execution compensation
A lightweight model, trained on a small amount of real deployment data, corrects tracking errors and dynamics mismatch on the specific robot.
Sub-1 cm manipulation accuracy on a Unitree G1
Hierarchical coordination reasoning
The high-level policy reasons over real-time low-level feedback and decides when and how to invoke skills, so execution stays consistent under latency.
Low-latency inference
Why a robot cannot simply move faster.
At an end-effector speed of 3 m/s, 10 ms of delay is about 3 cm of movement — enough to miss a grasp. Mind-1 rebuilds the whole loop from camera exposure to robot action.
- 82 ms → 32 ms
- Model inference61% lower on average
- 153 ms → 77.5 ms
- Observation to actionEnd-to-end latency
- Under 5 ms
- Sensor synchronisationAverage temporal alignment error
What comes next
From demonstration to deployment.
Reaching human-level task cycle time does not mean every challenge of real-world deployment has been solved.
Long-term reliability, failure recovery, safety, integration with existing workflows and deployment economics all need to be validated in real operating environments. That is where Mind-1 is going next, and it is what our feasibility studies and proofs of concept are for. How we work with customers →
Questions
Mind-1 and MindOn's technology
What is Mind-1?
Mind-1 is MindOn's Physical AI model, introduced in September 2026. It is designed to bring robot task execution to human-level speed: across the logistics and everyday manipulation tasks MindOn has evaluated, it reaches task cycle times comparable to a person, and on packaging and parcel sorting it is faster.
Which robots does Mind-1 run on?
The same model has been deployed on three robot embodiments: a fixed dual-arm robot, a wheeled dual-arm robot and a humanoid robot. MindOn separates a high-level intelligence layer from a low-level control layer, so the model is not tied to one robot body.
Does MindOn train on teleoperation data?
No. The model is trained on human-centric data: people performing tasks at natural working speed, captured with egocentric, handheld and whole-body devices. A small amount of real robot deployment data is used only for a lightweight compensation model that corrects tracking errors on the specific robot.
How does MindOn measure robot speed?
By end-to-end task cycle time: the time to complete the entire task from start to finish, not maximum joint velocity or isolated motion segments. For example, packaging takes Mind-1 about 3 seconds, against about 5 seconds for a person.
What is the difference between Mind-0 and Mind-1?
Mind-0, released in June 2026, showed that one model trained on human-centric data can run both humanoid and dual-arm robots. Mind-1 adds human-level speed: folding went from 80 seconds to 25 seconds per item, and average model inference latency fell from 82 ms to 32 ms.
What is not yet proven?
Human-level cycle time does not mean every deployment challenge is solved. Long-term reliability, failure recovery, safety, integration with existing workflows and deployment economics still need to be validated in real operating environments, which is what MindOn's pilots are for.
Could this fit your workflow?
Tell us about the task. We will come back with what a first feasibility assessment could look like.