Powered by elizaOSSimulation-first embodied data platform
Teleoperate robots.
From your browser.
No hardware. No expertise. Control simulated robot arms like a 3D web game, generate VLA-ready trajectory data, and own its provenance on Robinhood.
Human IK · Auto patrol ·Fly connectome
01 / Watch it live
A fruit fly's wiring,
driving the arm.
This is the Fly controller end-to-end. We take a visual crop of MaleCNS — the complete male fruit-fly nervous system mapped by Google Research and HHMI Janelia — wire the scene camera into its photoreceptors, and read descending neurons back out as end-effector motion. Collision or success fires a dopamine-style teach event.
No training. No policy. Fixed biological wiring — logged as controller=flycns in every episode.
SCENE CAMERA → PHOTORECEPTORS → DESCENDING-NEURON READOUT → EE_DELTA · PPL101 / PAM TEACH EVENTS ON COLLISION OR SUCCESS.
02 / The bottleneck
Physical AI is starving for demonstration data.
Language models scaled on internet text. Robot policies can't — every trajectory has to be demonstrated by a human. Today that means expensive robot rigs, slow lab pipelines, and datasets too small and too uniform to generalize. phyzical removes the hardware from the loop: anyone, anywhere, contributes manipulation demonstrations through a browser tab.
A browser is the only hardware a contributor needs.
Joint states, end-effector pose and object poses, every frame.
Isaac Sim re-rendering and domain randomization multiply every demo.
03 / Protocol
From browser tab
to training data.
The full loop takes minutes, not months. Every step below ships in the MVP — no future tense, no vaporware.
Pick a task
Browse the task board — sorting, pick-and-place, drawer opening. Each task shows difficulty, estimated time and reward before you start.
Teleoperate in-browser
Three inputs feed the same WebSocket: Human drags the end-effector and IK solves the joints; Auto runs scripted patrol; Fly routes the scene camera into R1–R6/R8 photoreceptors and maps descending neurons to relative motion.
Trajectory auto-recorded
The client samples joint positions, EE pose and object poses at 30–60Hz. Fly episodes additionally log controller, fly_stim, dn_readout and teach_events — same schema, different actions source.
Submit, verify, own
One click submits the episode. Controller-aware QC runs — smoothness and success for Human; collision rate and visual habituation for Fly — then a Data ID is anchored on Robinhood.
04 / Controllers
Three ways to move
the same arm.
One episode schema, one WebSocket, three sources of actions. Humans produce the pre-training core; Fly produces differentiated rollouts and control baselines.
Human
Mouse-drag end-effector with real-time IK. Produces demonstrations with intent, timing and strategy — the primary pre-training data.
Auto
Scripted patrol routines. Coverage sweeps, workspace baselines and regression references for every task scene.
Fly
A MaleCNS visual crop wired to the scene camera: photoreceptors in, descending-neuron readout out, dopamine-style teach events on collision or success. Fixed biological wiring, experimental readout — tagged controller=flycns.
FLY MODE IS A MAPPED CONNECTOME CONTROLLER. IT IS NOT A TRAINED POLICY AND NOT A CLAIM OF ANIMAL-LEVEL DEXTERITY.
05 / Task board
Manipulation tasks,
gamified.
Block Sorting
Sort colored blocks into matching zones. Human drives the strategy; Fly can cut in as a collision-avoid reflex layer.
Pick & Place
Grasp objects of varied geometry and place them on randomized targets.
Drawer Opening
Articulated manipulation — approach, grasp handle, pull along constrained axis.
Stacking Tower
Stack blocks into a stable tower. Physics-checked stability over 3 seconds.
Keep Target Centered
Visual servoing — hold a moving color patch in the camera center. Compound eyes were built for exactly this.
Approach & Stop
Drive toward a color block, stop on contact. Aversive teach event fires on overshoot or collision.
Push Block to Zone
Forward + turn only, no grasping. Push the block into the goal zone using DN drive/turn readout.
Cable Routing
Deformable object manipulation. Route a cable through fixed guides.
Dual-Arm Handover
Coordinated two-arm transfer with grasp re-orientation mid-air.
Every completed episode enters the network queue: quality checks, augmentation, then on-chain registration. This preview shows the pipeline event format shipping with the MVP — including the controller tag.
06 / Data pipeline
One human demo.
Hundreds of training samples.
Capture
Unity client records joint states, EE pose, object poses and user actions at fixed physics ticks. Buffered locally, submitted in one shot.
Quality check
Controller-aware QC. Human episodes: success criteria, smoothness, duration bounds, replay-attack filters. Fly episodes: collision rate, drop events and visual habituation.
Augmentation
Accepted episodes are re-rendered in Isaac Sim with photorealistic lighting and domain randomization — one demo becomes hundreds of training samples.
Dataset
Versioned, filterable datasets exported in VLA / imitation-learning-ready formats for labs and model teams.
1{2 "episode_id": "EP_7f3a2c",3 "task_id": "block_sorting_01",4 "contributor": "0x94e1…c7f2",5 "sim": { "engine": "unity_webgl", "physics_hz": 60 },6 "duration_s": 45.2,7 "success": true,8 "trajectory": [9 {10 "t": 0.033,11 "joint_pos": [0.12, -0.87, 1.42, -0.55, 0.00, 0.31],12 "ee_pose": { "xyz": [0.42, 0.31, 0.87], "quat": [0, 0, 0.71, 0.71] },13 "gripper": 0.04,14 "object_poses": { "block_red_01": { "xyz": [0.40, 0.02, 0.90] } },15 "action": { "type": "ee_delta", "cmd": [0.01, 0.00, -0.02] }16 }17 ],18 "controller": "flycns_v1",19 "fly": {20 "graph": "malecns_v1_visual_crop",21 "photoreceptors": { "r1r6": 3335, "r8": 811 },22 "readout": ["DNp20", "DNpe017"],23 "teach_events": [24 { "t": 12.4, "cell": "PPL101", "kind": "aversive", "ms": 200 }25 ]26 },27 "provenance": {28 "data_id": "phyz:robinhood:0x3d9f…a1b4",29 "content_hash": "sha256:e3b0c442…",30 "anchored_at": "block 24911072"31 }32}07 / Ecosystem
The fuel station for
agent robot stacks.
Agent frameworks like elizaOS now ship full robot stacks — simulation, RL training, real hardware bridges. What they need is human demonstration data. That's what phyzical produces.
Schema-aligned trajectories
phyzical episodes — joint_pos, ee_pose, actions and object poses at 30–60Hz — map directly onto eliza_robot's trajectory database. One converter, no re-collection.
Profile-matched tasks
Tasks are pinned to real robot profiles (RobotProfileId) — the same URDF/MJCF morphologies the elizaOS robot stack simulates and deploys, from arms to humanoids like AiNex and ASIMOV-1. The stack gains a FlyController profile alongside.
Warm-start training
Human demonstrations bootstrap imitation learning and offline RL — feeding continual-RL and text-conditioned policies instead of exploring from zero.
Browser demo in → trainable episode out
phyzical fuels embodied agents. elizaOS-ready by design.
08 / Provenance
Your data.
Verifiably yours.
Crowdsourced data is worthless without trust in where it came from. phyzical anchors each episode's identity, hash and contributor on Robinhood — so labs can audit datasets, and contributors keep a claim on the value they create.
Data ID on Robinhood
Every accepted episode gets a unique Data ID with its content hash and controller tag anchored on Robinhood — verifiable by anyone, filterable by controller.
Ownership follows the contributor
The wallet that produced the demonstration is written into the record. Provenance isn't a promise in a database — it's public state.
Foundation for a data economy
Verifiable origin is what makes data tradable. Registered episodes are the primitive for future licensing, incentives and marketplace flows.
09 / Roadmap
Shipping in phases.
Web Teleop Core
- Unity WebGL scene + URDF arm
- Mouse-drag end-effector + IK
- Task success detection
MVP Loop
- Next.js shell + wallet connect
- WebSocket state sync
- Task board + progress
Data & Chain
- Trajectory submit + QC pipeline
- Data ID registry on Robinhood
- Points rewards live
Scale & Augment
- Isaac Sim photoreal augmentation
- Domain randomization at scale
- Dataset access for labs
- Fly controller · visual crop → hybrid safety layer