Introducing P37 Neuro: Architecture and Cross-Embodiment Intelligence

Why embodiment is part of the model

A policy cannot treat every robot as the same actuator layout. P37 Neuro represents links, joints, limits, control modes, end effectors, sensor inventory, morphology, and action-space metadata as explicit machine data.

This makes embodiment a condition of learning rather than a hidden assumption. The long-term research question is whether useful concepts such as reach, grasp, locomotion, and recovery can transfer while low-level execution remains specific to a body.

A hierarchy from intent to actuation

P37 separates a slower high-level policy from a fast, body-specific low-level policy. The high-level layer tracks task progress, selects subgoals, coordinates behaviors, and plans recovery. It should express intent without issuing raw torque.

The low-level layer realizes that intent through locomotion, balance, contact handling, trajectories, and joint commands. This separation provides an explicit boundary for morphology randomization, reinforcement learning, and sim-to-real evaluation.

Episodes, memory, and recovery

The canonical episode contract records identity and provenance, embodiment, task context, observations, actions, outcomes, interventions, failures, and timing. Robot-specific adapters may differ, but stored learning data should not fragment by manufacturer.

Successful and unsuccessful attempts are both retained. Long-context memory is intended to expose prior failures, changed payloads, blocked joints, and task progress to later decisions instead of discarding them after one control window.

Safety and current status

Learned policies remain behind a deterministic runtime responsible for position, velocity, effort, freshness, workspace, watchdog, and health constraints. Hardware integrations must still supply independent emergency stops and controller-level safety.

The repository implements platform paths across data, training, simulation, runtime, artifact, and rollout infrastructure. Large-scale capability training, held-out embodiment evaluation, multi-robot physical validation, long-duration recovery, and production fleet qualification remain active work.

• Platform implementation is not equivalent to qualified physical capability.
• Simulation provides scale; hardware validation supplies deployment evidence.
• Every promoted capability needs held-out bodies, tasks, and environments.