Reflex

Deploy any open vision-language-action policy to a real robot in one command.

What Reflex is

Reflex is the deployment layer between a trained robot brain and a real robot. It takes a vision-language-action policy — pi0, pi0.5, SmolVLA, NVIDIA’s GR00T — from HuggingFace and runs it on an edge GPU like a Jetson Orin or Thor in a single command, with verified machine-precision numerical parity to PyTorch.

pip install reflex-vla and reflex go --model smolvla-base is the entire surface area to get started.

One command from model to robot

reflex go --model <hf_id> probes hardware, downloads weights, exports ONNX, builds a TensorRT engine, and starts an HTTP server. No editing configs, no separate reflex export step.

Verified parity, not 'close enough'

Every supported model exports with cos = +1.000000 numerical parity to the PyTorch reference. Most “deploy your model” tools settle for “close enough” and quietly diverge in production. Reflex does not.

Edge-first

Designed for Jetson Orin Nano / AGX / Thor and desktop NVIDIA GPUs (RTX 30 / 40 / 50). Cloud A10G / A100 / H100 also supported for benchmarking and validation runs.

Composable wedges

Safety clamping, action-chunk correction, deadline guards, cloud fallback, traces, and ROS2 transport — every wedge is a flag on reflex serve. Compose only what you need.

Verified across the four major open VLAs

Model	Parameters	First-action max_abs vs PyTorch
SmolVLA	450M	5.96 × 10⁻⁷
pi0	3.5B	2.09 × 10⁻⁷
pi0.5	3.62B	2.38 × 10⁻⁷
GR00T N1.6	3.29B	8.34 × 10⁻⁷

All four pass at machine precision on shared seeded inputs. See Verified parity for the full ledger.

Status

v0.7 — source-available under BSL 1.1, the same license HashiCorp, MongoDB, and Sentry use. Auto-converts to Apache 2.0 in 2030. Active development. We’re looking for the first 20 robotics teams actually deploying this — your feedback shapes v0.8.