PHASE 0 · PROJECT THESIS
Make robot-policy transfer a testable claim.
We help developers using low-cost, LeRobot-compatible manipulation arms determine whether a published policy can run on their hardware—and understand why when it cannot.
01
The first community
Our first community is practitioners actively attempting to reproduce or adapt publicly shared manipulation policies on SO-100, SO-101, and closely compatible low-cost LeRobot hardware.
It includes independent builders, robotics students, educators, open-source maintainers, small laboratories, and early robotics teams. They already share code, policies, datasets, calibration guidance, and failure reports through GitHub, Hugging Face, Discord, papers, and classrooms.
02
The painful workflow
A practitioner finds a promising policy, then reconstructs the conditions under which it worked: exact revision, robot configuration, calibration, cameras, observation and action conventions, dependencies, control timing, and evaluation protocol. Much of this is implicit or distributed across documentation and issue threads.
“Can this policy run on my setup, and what must be true for its reported result to mean anything here?”
That is the decision we intend to make faster and more defensible.
03
Why this wedge is credible
LeRobot contributors, arm owners, students, maintainers, educators, and small labs are reachable in existing technical communities.
Public reports document calibration, gripper-direction, hardware-interface, and configuration failures.
Time to first valid trial, successful trials, interventions, and independent reproductions can be counted.
LeRobot supplies interfaces and dataset conventions; GitHub and Hugging Face already host the artifacts.
Early validation can use contributor-operated hardware and local execution, with trust levels stated explicitly.
A proven record can later cover more arms, simulations, physical agents, and accredited testing.
04
Deliberate boundaries
During this discovery cycle, Known Robot is not a robot manufacturer, policy marketplace, safety certifier, simulation engine, model host, or universal robotics standard.
We are studying one decision: whether a third party has enough evidence to attempt a policy transfer on low-cost manipulation hardware.
05
Falsifiable Phase 1 hypotheses
- Practitioners spend meaningful time reconstructing configuration before a valid trial.
- Calibration and sensor configuration are at least as important as model metadata in transfer failures.
- Success rates are frequently not comparable across repositories without additional protocol details.
- Practitioners will share non-confidential failure evidence when attribution and scope are clear.
- At least five qualified teams will commit an artifact, evaluation, cadence, publishable result, or introduction.
We will change or abandon the wedge if 25 conversations do not support a repeated workflow with measurable cost.
HELP TEST THE THESIS