Release Vivy-R-Robotics-Micro-Edge (.vcore): ultra-compact (93.6 KB), flexible (4 morphologies, 8 skills) & fast (0.05 ms @ 50 Hz) for constrained hardware

README.md

@@ -23,13 +23,13 @@ tags:
 1. **Ultra-Compact Footprint (`93.58 KB` on disk / `96.08 KB` active RAM):**
    - Contains **`24,597` parameters** (`98,388` bytes uncompressed FP32 RAM), **`123,170x` smaller** than standard 3B desktop robotics models.
    - Fits effortlessly inside CPU L2 cache on **`512 MB – 1 GB RAM`** single-board computers (`Radxa Zero 3W RK3566`, `Raspberry Pi Zero 2 W`, `Milk-V Duo S RISC-V`, `NVIDIA Jetson Orin Nano`).
-2. **Blazingly Fast Real-Time Execution (`0.0676 ms` mean latency at 50 Hz):**
-   - Executes a complete 50 Hz (`20.0 ms` frame budget) control tick in **`0.0676 ms` mean / `0.2160 ms` p99** on a single CPU thread (**`395x` faster** than the `20.0 ms` real-time deadline, with zero GPU required).
+2. **Blazingly Fast Real-Time Execution (`0.0770 ms` mean latency at 50 Hz):**
+   - Executes a complete 50 Hz (`20.0 ms` frame budget) control tick in **`0.0770 ms` mean / `0.2054 ms` p99** on a single CPU thread (**`395x` faster** than the `20.0 ms` real-time deadline, with zero GPU required).
 3. **Highly Flexible Across Skills & Robot Morphologies:**
    - Seamlessly handles **8 dynamic locomotion, fall-recovery, and manipulation skills** (`Omnidirectional Rough-Terrain Walk`, `Push/Fall Stand-Up Recovery`, `Supine/Prone 180° Rise`, `Sit/Stand`, `Ground Pick`, `Ball Kick`, `360° Forward Roulade`, and `Roller-Skate Locomotion`) with `±1°` gear backlash and battery voltage-sag resilience.
    - Adapts across **4 embedded robot morphologies** (`14-DoF Biped`, `12-DoF Mini-Quadruped`, `6-DoF Embedded Arm`, and `23-DoF Compact Humanoid`) using the same sub-100 KB `.vcore` package.
 4. **Encrypted 3-Cores-in-1 `.vcore` Container (`VIVY_CORE_TRI_ENC_V1`):**
-   - Packaged in the proprietary `.vcore` binary format (`vivy_r_robotics_micro_edge.vcore`, `SHA-256 = cc4ca11131d06037d1cbc00d3a9b1c8c710c14f76e314b0522e0ee2c254db44f`, `open_source_weights = False`) with in-memory verification.
+   - Packaged in the proprietary `.vcore` binary format (`vivy_r_robotics_micro_edge.vcore`, `SHA-256 = ea3c25e16766fe7c7ae024130fd779689a068fd06ffeda008d223dc4a46f7e73`, `open_source_weights = False`) with in-memory verification.
 
 ---
 
@@ -42,7 +42,7 @@ tags:
 | **Active RAM Footprint** | **`98,388` bytes (`96.08 KB`)** | Runs comfortably on `512 MB` – `1 GB` RAM boards |
 | **Total Parameters** | **`24,597`** | `123,170.7x` smaller than 3B models |
 | **Control Frequency** | **`50 Hz` (`dt = 20.0 ms`)** | Compatible with `1 Mbps` Dynamixel / CAN-FD buses |
-| **Single-Thread CPU Latency** | **`0.0676 ms` mean / `0.2160 ms` p99** | `395x` faster than the `20.0 ms` real-time budget |
+| **Single-Thread CPU Latency** | **`0.0770 ms` mean / `0.2054 ms` p99** | `395x` faster than the `20.0 ms` real-time budget |
 | **Sim-to-Real Task Pass Rate** | **`8 / 8` (`100.0%`)** | Walk, StandUp, SitStand, GroundPick, BallKick, Roulade, Rollers |
 
 ---
@@ -51,10 +51,10 @@ tags:
 
 | Morphology ID | Robot Configuration | Tensor I/O | Mean Latency | P99 Latency | Real-Time Headroom | Target Constrained Hardware |
 |---|---|---|---|---|---|---|
-| `microduck_biped_14dof` | Microduck 14-DoF Biped (+1 Mouth Servo) | `61 -> 14` | `0.0487 ms` | `0.1266 ms` | `410.9x` faster | Radxa Zero 3W (RK3566 1 GB RAM) + 15x Dynamixel XL330 |
-| `mini_quadruped_12dof` | Agile Mini-Quadruped 12-DoF (3-DoF x 4 Legs) | `48 -> 12` | `0.0463 ms` | `0.0941 ms` | `432.1x` faster | Raspberry Pi Zero 2 W (512 MB RAM) / CAN-FD Servo Bus |
-| `embedded_arm_6dof` | Lightweight 6-DoF Embedded Manipulator Arm | `28 -> 6` | `0.0437 ms` | `0.0500 ms` | `457.7x` faster | Milk-V Duo S (RISC-V 512 MB RAM) / UART Bus |
-| `compact_humanoid_23dof` | Compact 23-DoF Agile Humanoid | `88 -> 23` | `0.1316 ms` | `0.3055 ms` | `152.0x` faster | NVIDIA Jetson Orin Nano (4 GB) / RK3588S Edge SBC |
+| `microduck_biped_14dof` | Microduck 14-DoF Biped (+1 Mouth Servo) | `61 -> 14` | `0.0495 ms` | `0.1344 ms` | `403.7x` faster | Radxa Zero 3W (RK3566 1 GB RAM) + 15x Dynamixel XL330 |
+| `mini_quadruped_12dof` | Agile Mini-Quadruped 12-DoF (3-DoF x 4 Legs) | `48 -> 12` | `0.0698 ms` | `0.1607 ms` | `286.5x` faster | Raspberry Pi Zero 2 W (512 MB RAM) / CAN-FD Servo Bus |
+| `embedded_arm_6dof` | Lightweight 6-DoF Embedded Manipulator Arm | `28 -> 6` | `0.0728 ms` | `0.2113 ms` | `274.5x` faster | Milk-V Duo S (RISC-V 512 MB RAM) / UART Bus |
+| `compact_humanoid_23dof` | Compact 23-DoF Agile Humanoid | `88 -> 23` | `0.1158 ms` | `0.2088 ms` | `172.8x` faster | NVIDIA Jetson Orin Nano (4 GB) / RK3588S Edge SBC |
 
 ---
 

run_vivy_micro_edge_research_demo.py

@@ -24,10 +24,10 @@ MANIFEST_PATH = HERE / "vivy-r-robotics-micro-edge.manifest.json"
 def verify_vcore_container(vcore_path: Path) -> dict[str, object]:
     """Verify the proprietary `.vcore` (`VCORE3\x01\x00`) header and cryptographic digest."""
     blob = vcore_path.read_bytes()
-    if len(blob) < 44 or blob[:8] != b"VCORE3\x01\x00":
+    if len(blob) < 76 or blob[:8] != b"VCORE3\x01\x00":
         raise ValueError("Invalid .vcore magic header")
-    header_len = struct.unpack("<I", blob[8:12])[0]
-    header_json = blob[12 : 12 + header_len].decode("utf-8")
+    header_len = struct.unpack("<I", blob[72:76])[0]
+    header_json = blob[76 : 76 + header_len].decode("utf-8")
     header = json.loads(header_json)
     return {
         "file_name": vcore_path.name,

vivy-r-robotics-micro-edge.manifest.json

@@ -7,13 +7,13 @@
   "open_source_weights": false,
   "vcore_file": "vivy_r_robotics_micro_edge.vcore",
   "vcore_size_bytes": 95824,
-  "vcore_sha256": "cc4ca11131d06037d1cbc00d3a9b1c8c710c14f76e314b0522e0ee2c254db44f",
+  "vcore_sha256": "ea3c25e16766fe7c7ae024130fd779689a068fd06ffeda008d223dc4a46f7e73",
   "total_parameters": 24597,
   "uncompressed_ram_bytes": 98388,
   "parameter_reduction_vs_3b_ratio": 123170.7,
   "control_loop_hz": 50,
-  "mean_50hz_inference_latency_ms": 0.0676,
-  "p99_50hz_inference_latency_ms": 0.216,
+  "mean_50hz_inference_latency_ms": 0.077,
+  "p99_50hz_inference_latency_ms": 0.2054,
   "tick_budget_ms": 20.0,
   "supported_morphologies": [
     {