title: README
emoji: 🤖
colorFrom: green
colorTo: gray
sdk: static
pinned: false
ZenEdgeX
Proprio: exact inference for robot policies. Proprio runs unmodified vision-language-action checkpoints faster and verifies every action chunk bit for bit against the original model before serving it. If the actions are not identical, it refuses to deploy. Every run writes a Bitwise Certificate anyone can rerun.
New in 0.1.2: a drop-in for LeRobot async inference. proprio lerobot-server takes the same
flags as LeRobot's policy server; the robot client does not change. Against LeRobot's own server on
an RTX 4090, same observations and seeds: SmolVLA 20/20 action chunks identical, about 330 ms to
52 ms per chunk; pi0.5 20/20 identical. Guide: https://github.com/zenedgex/proprio-releases/blob/main/LEROBOT.md
Model tick, eager against Proprio, all ticks identical (RTX 4090):
| Policy | Eager | Proprio | |
|---|---|---|---|
| lerobot/smolvla_base | 232 ms | 37–40 ms | 5.9–6.3x |
| lerobot/pi05_base (bf16) | 257 ms | 97 ms | 2.7x |
| nvidia/GR00T-N1.7-3B | 170 ms | 64 ms | 2.7x |
| openvla/openvla-7b | 209 ms | 145 ms (H100: 88 ms) | 1.4x (H100: 2.7x) |
| 16 SmolVLA robots per tick | 240 ms | 19.3 ms per robot | 12.5x, every row identical to the robot alone |
- Evaluation wheels, free, no key (Linux x86_64 and aarch64): https://github.com/zenedgex/proprio-releases
- The certificates and server comparisons behind these numbers: the
proprio-certificatesdataset in this organization - Results: https://zenedgex.in/proprio/
- Serving: LeRobot async-inference server, openpi-compatible websocket server,
proprio doctorto check a machine first
Patent pending. Contact: hello@zenedgex.in