P37 NEURO
A shared intelligence program for different robot bodies.
P37 Neuro is an open-source research and engineering system that represents embodiment explicitly and separates task-level intelligence, body-specific control, and deterministic safety.
Objective and system boundary
P37 Neuro is designed around a difficult long-term objective: a new robot body or task should not automatically require an entirely new model. It treats perception, task context, embodiment, action, memory, and evaluation as explicit system boundaries.
The program is not a thin ROS wrapper or an unrestricted neural controller. Learned policy output remains behind deterministic actuator constraints and hardware-specific safety systems.
High-level and low-level policies
A slower high-level policy is intended to track task progress, select subgoals, plan recovery, and express embodiment-neutral or weakly body-dependent intent. A faster low-level policy converts that intent into body-specific motion.
Robot specifications encode links, joints, limits, actuators, end effectors, sensors, morphology, and action-space metadata. Canonical episode data records observations, actions, outcomes, interventions, failures, and provenance across machines.
Learning, memory, and simulation
The repository includes behavior-cloning and PPO primitives, recovery and disturbance curricula, a variable-morphology recurrent baseline, MuJoCo execution, and an Isaac Lab locomotion task. Simulation adapters remain optional so the core contracts can be tested without a full simulator.
The research roadmap uses successful and unsuccessful attempts as learning context. Long-horizon recovery and broad transfer are capability goals that still require qualification.
Runtime and developer tools
Implemented platform components include URDF, MJCF, and OpenUSD ingestion; canonical episode contracts; content-addressed artifacts; signed releases; a C++20 actuator-safety runtime; an optional ROS 2 bridge; and an optional TensorRT recurrent inference backend.
The CLI supports robot manifests, validation, qualification evidence, rollout plans, status, and rollback artifacts. It does not bypass controller or site approvals.
Current validation status
The public repository describes the executable platform from its E0 foundations through E6 deployment infrastructure as implemented. It explicitly distinguishes platform completion from capability completion.
Large-scale training, held-out embodiment evaluation, target-GPU profiling, multi-robot physical validation, long-duration recovery, and live fleet qualification remain in progress. No claim of general-purpose physical capability should be inferred from the platform alone.
Common questions
What is P37 Neuro?
P37 Neuro is Ontos World’s research and engineering program for cross-embodiment robot policies that connect perception, task context, embodiment, memory, and action.
How does cross-embodiment learning work?
The architecture represents each robot body as structured data, keeps a common learning contract, and separates transferable task intent from body-specific low-level control.
Which systems can P37 Neuro integrate with today?
The repository implements robot-description ingestion, MuJoCo execution, an Isaac Lab task, and optional ROS 2 and TensorRT boundaries. Physical compatibility still requires robot-specific integration and qualification.