01
Real task
RGB-DPoseTask stageOutcome
Manort Task Data
Real work becomes a place a robot can learn.
We are developing a workflow that captures how people perform physical tasks—or starts with data a company already controls—and turns it into structured demonstrations, paired simulation, and training inputs for a defined robot and workflow.
In development at Manort
02
Physics twin
GeometryObjectsPhysicsVariation
03
Real-to-Sim Task Pack
ObservationActionContextRobot fitEvaluation
Structured for one defined task and compatible target system.
Start with the work
Capture a task with us. Or bring what you already have.
Every project begins with a specific task, environment, and target robot. The input can come from a new capture project or suitable material already collected by the customer.
Capture the signals the task needs.
We define the task and select the viewpoints and sensing needed to record movement, objects, environment, and outcome for the intended training use.
Plan a capture ↗Start with data you already control.
Video, depth recordings, teleoperation sessions, robot logs, CAD, or site layouts can provide a starting point. We assess rights, timing, calibration, task context, and any information that is still missing.
Assess existing data ↗Real to sim
One task. A connected learning workflow.
Human demonstrations become more useful when the action, environment, and outcome are connected. Manort pairs real evidence with a physics-based version of the task.
Define the work
Agree on the site, objects, target robot, starting conditions, and a measurable definition of success.
Structure the demonstration
Organize usable observations, movement or action signals, task stages, contact events, and outcomes into synchronized episodes.
Pair real with simulation
Recreate the relevant environment and prepare compatible actions, controlled variations, and failure cases in Manort Simulator Engine.
Validate the task pack
Test held-out situations and refine the data before supervised work with the physical robot.
More than video
The context around every action.
Depending on the source material and project scope, a Real-to-Sim Task Pack can connect the demonstration to the place, robot, and evaluation that give it meaning.
- Context
- Scene geometry, task objects, instructions, initial conditions, and the intended goal.
- Demonstration
- Time-aligned observations, movement or action signals, task stages, contacts, and outcomes.
- Simulation
- A paired task environment with relevant geometry, physical properties, and controlled variations.
- Robot fit
- A project-specific mapping for the agreed embodiment, end effector, sensors, and control interface.
- Evaluation
- Quality documentation, training and validation splits, and a benchmark defined around the task.
01
02
03
04
05
For Milo and beyond
Built with Milo. Prepared for compatible systems.
A company does not need to use our robot to explore this workflow. We can assess selected customer robots by reviewing their kinematics, end effector, sensors, control interface, and the task they need to perform.
Compatibility and deliverables are agreed before a project begins.
Discuss your robotResponsible by design
Clear rights from source to delivery.
Each project defines who may provide the source material, how it may be used, what the customer receives, and how long identifiable material is retained. Customer data is considered only when the customer has the necessary rights and permissions to share it.
Where people appear, collection planning covers the appropriate legal basis, participant and site permissions, access controls, and data minimization. Identifiable recordings can be separated from derived task data, with redaction or de-identification used where appropriate.
Questions
Before the first task.
Scope, compatibility, rights, and evaluation are defined around the project.
Do we need to use Milo?
No. The workflow is being developed with Milo, but we can assess selected customer robots. Suitability depends on the robot, end effector, control interface, available data, and task.
Can we provide data we already collected?
Yes, subject to review. Useful inputs may include video, depth data, teleoperation episodes, robot logs, site layouts, CAD, and task documentation. We first check rights, quality, calibration, timing, and whether the material contains enough information for the intended use.
Is ordinary video enough to train a robot?
Video can describe a task and its environment. Control training often also needs poses, actions, gripper state, object state, timing, or robot feedback. We identify those gaps before proposing additional capture.
What can a project deliver?
Depending on scope, a delivery can include structured task episodes, annotations, a paired simulation environment, robot-specific mappings, quality documentation, and a defined evaluation set.
How is customer and participant data handled?
Each project defines source rights, participant and site permissions, access, retention, redaction, permitted model use, and what the customer receives before collection or conversion begins.
Does simulation guarantee real-world performance?
No. Simulation moves early iteration away from a live site and exposes a system to controlled variations. Performance and safety still require supervised validation with the target robot in the real environment.
Start with one task
Show us the work you want a robot to learn.
Tell us about the task, site, target robot, and data you already have. We can assess a focused Real-to-Sim pilot and define what must be captured, converted, simulated, and validated.