Introducing Manifold Intelligence: Building Intelligence for the Physical World
We are building a scalable path from human demonstrations to learned robot behavior on existing industrial arms.
By Veer Patel, Founder, Manifold Intelligence
Industrial robots are powerful, precise, and reliable when the world around them is carefully structured. The difficulty begins when objects shift, materials vary, or a task changes often enough that reprogramming becomes expensive.
Manifold Intelligence is being built around a long-term goal: creating general-purpose intelligence for the physical world.
We are starting with a specific thesis - robots should be able to learn useful manipulation behavior from scalable human demonstrations and transfer those skills to existing industrial arms.
A different starting point for robot learning
Many robot-learning systems collect demonstrations directly on physical robots through teleoperation or engineered setups. That approach is important, but it ties collection to robot access, trained operators, and time on the machines themselves.
Manifold is exploring a complementary approach. We use a wearable demonstration-capture system that allows people to perform manipulation tasks naturally while recording the visual context, movement, and gripper state needed for robot-learning research.
This means demonstrations can be collected without occupying an industrial robot, across different people, sessions, and environments. It also captures variation in how people approach the same physical task - variation that a small set of scripted robot trajectories may not contain.
Human motion does not transfer directly to a robot. Differences in embodiment, coordinate systems, control, timing, and safety still need to be solved. Our work is focused on building that bridge.
What we have built so far
Manifold has developed an engineering loop that takes demonstrations from raw recordings to model evaluation. It supports collection quality control, segmentation into manipulation episodes, trajectory and gripper-state extraction, structured dataset generation, policy training, source-grouped offline evaluation, and live-camera robot-off inference.
Our current engineering dataset contains approximately 26.4 hours of demonstrations from 342 source recordings across four task families.
We began with single-hand demonstrations to establish the collection, processing, training, and evaluation loop. We have also collected a small exploratory bimanual pilot dataset as an early step toward understanding synchronized two-hand behavior.
The pilot is not yet evidence of a trained bimanual policy or physical bimanual execution. It is helping us define what coordinated observations, actions, and eventual two-arm control will require.
One lesson from the data
Building the pipeline has repeatedly shown us that a model is only as trustworthy as the experiment behind it.
Early in development, long source recordings were segmented into multiple task episodes. Randomly placing some episodes in training and others in validation would create two different numerical sets, but not a genuinely independent evaluation.
Episodes from the same source recording can share lighting, camera placement, background, operator, object arrangement, and neighboring moments from the same attempt. A model can appear to generalize while being tested on conditions extremely similar to those it already saw.
We changed the evaluation process so episodes from one source recording remain together on the same side of the split.
It is a small example of a principle guiding Manifold: trustworthy progress sometimes requires making the evaluation harder. A higher metric is not automatically a better result if the experiment behind it is weak.
From offline predictions to physical execution
Offline evaluation is necessary, but physical intelligence cannot be established from validation loss alone. During execution, the policy's own actions create the next observation. A small prediction error can move the system into an unfamiliar state, where later errors may compound.
We currently use robot-off inference to study this gap safely. The model receives live camera observations and generates predicted actions, but no robot receives those commands. This allows us to inspect what the model sees, how predictions change with the scene, and whether the resulting action traces are physically plausible.
Our next technical milestone is repeatable robot-off comparison under a new camera and new collection conditions, followed by measured single-arm execution in a controlled safety environment.
When we reach physical execution, we intend to measure task success across repeated trials, human intervention, safety-stop events, execution time, response to scene changes, and recovery after unsuccessful actions.
Why existing industrial arms
Factories already contain capable robot hardware. The challenge is often the cost and difficulty of adapting those systems to variable tasks, frequent changeovers, or new object configurations.
Our long-term aim is to build an intelligence layer that works with existing industrial platforms rather than requiring every customer to adopt an entirely new robot embodiment.
If successful, this could reduce the task-specific programming and demonstration effort required for new manipulation workflows, while retaining the industrial hardware, support, and safety infrastructure manufacturers already understand.
What comes next
• Test policies under genuinely new camera and collection conditions.
• Identify industrial workflows where variability makes conventional automation expensive.
• Establish supervised access to reliable robot hardware.
• Measure single-arm execution before progressing to coordinated bimanual control.
Manifold is early, but it has moved from an idea to a working demonstration collection, processing, training, and robot-off evaluation system. The infrastructure now allows us to pursue the next stage with evidence rather than assumption.
We are looking for partners
We would like to speak with:
• Manufacturers with repetitive manipulation workflows that remain difficult to automate.
• System integrators working around variation, changeovers, or high programming costs.
• Robotics engineers interested in perception-to-execution systems.
• Testing and design partners with access to supervised robot environments or narrowly defined manipulation problems.
If this overlaps with your work, contact us at vepatel@manifoldpi.com.
General-purpose physical intelligence will not come from a model alone. It will require data, perception, learning, control, hardware, safety, and disciplined physical evaluation to work as one system.
That is what we are building at Manifold Intelligence.
We intend to earn the right to put intelligence into motion.
