Phyzicalorg/Phyzical_org

What it solves

Phyzical provides a way to collect large-scale human demonstration data for robot training without requiring physical hardware or specialized expertise. It allows users to teleoperate simulated robot arms via a web browser, creating high-frequency (30–60Hz) trajectory data that can be used to train Vision-Language-Action (VLA) models and imitation learning policies.

How it works

Users interact with a simulated environment in a browser using Unity WebGL, where they drag the robot's end-effector and inverse kinematics (IK) handles the joint movements. The platform records these sessions as episodes containing joint positions, end-effector poses, and object poses.

This data is then processed through a pipeline:

1. Collection: Browser-based teleoperation (Human), scripted patrols (Auto), or a mapped fruit-fly nervous system controller (Fly).
2. Conversion: A converter transforms these episodes into a SQLite trajectory_db format compatible with the elizaOS robotics stack.
3. Provenance: Episode data is anchored on-chain to track contributors and content hashes.

Who it’s for

• AI Researchers: Those training VLA models or using imitation learning/RL for embodied agents.
• Robot Data Contributors: Non-experts who can provide demonstrations via a browser.
• Agent Robot Stacks: Developers using elizaOS who need high-quality human demonstration data to warm-start their policies.

Highlights

• Zero-Install Frontend: Teleoperation happens entirely in the browser.
• elizaOS Integration: Direct compatibility with elizaOS trajectory_db for seamless training pipeline integration.
• Fly Controller: An experimental controller based on the MaleCNS fruit-fly connectome map to generate reactive, non-human trajectories.
• On-chain Provenance: Uses blockchain to verify the identity and source of training episodes.