rhoyn.

Research

Measurements behind low-latency, precise VR teleop. Each article links an MIT-licensed repo with the code and raw recordings.

Reproductions and corrections: hello@rhoyn.com

Teleop latency2 articles

  1. Low-latency eyes

    Robot camera to VR headset, measured against the operator’s own passthrough.

    Mean lag vs passthrough
    rhoyn, D435+0.1 ms
    XRoboToolkit, ZED Mini+34.8 ms
    Method
    One spinning LED, seen through passthrough and the robot camera in the same headset frame. No clock sync.
    Setup
    Unitree G1, stock head D435.

  2. Low-latency, accurate hands

    Operator’s hand moves; how long until the robot’s hand does.

    Hand latency, median
    rhoyn20.3 ms
    XRoboToolkit + ExtremControl44.5 ms
    XRoboToolkit + GR00T89.3 ms
    Method
    Two IMUs, one per hand, on one ESP32 at 1 kHz. Median of ten algorithms.

Walking policy benchmarks2 articles

  1. Walk under disturbances

    43 publicly released Unitree G1 walking policies on one interface, run on the same set of punched 3-minute tours.

    Best completion ~5 in 6 38 of 69 ranked rows finish at most 1 in 100
    Task
    18 waypoints, punches up to 500 N, arms moving at random. Scored on recovery.
    Scale
    125,952 runs, MuJoCo + PhysX.

  2. Walk slowly and precisely

    How 43 publicly released G1 walking policies follow the small commands teleoperators mostly give; 34 configurations scored.

    Tracks small commands gr00t_wbc at 0.74 m; about half the rest don’t move below 0.13–0.2 m/s, or creep
    Task
    One walk or turn command held 10 s, scored by movement along it.
    Scale
    4096 runs per policy, command type and simulator (MuJoCo, PhysX), plus a turn test on a real G1.