Avero Tech builds open, developer-first infrastructure — autonomous agents, real-time vision, and edge-optimized models — that let robots perceive, decide, and act without a leash back to the cloud.
We don't believe intelligence for the physical world should ship behind an API key. Every core model, from perception encoders to motion planners, is released with open weights, reproducible training recipes, and hardware schematics — so the engineer wiring a servo at 2am has the same access as our own research team.
Avero's framework is modular by design: swap a vision backbone, fine-tune a policy head, or drop in your own PLC driver without forking the stack. Robotics moves too fast for walled gardens.
Every Avero deployment is composed from these primitives — mix, match, and fine-tune per robot.
Goal-directed agents that plan, call tools, and recover from failure without a human in the loop for every decision.
agent.core → stableSub-frame object detection and scene segmentation tuned to run on embedded GPUs at the edge, not a data center.
vision.rt → stableQuantized, distilled language models that reason on-device — no round trip to the cloud when milliseconds matter.
edge.llm → betaInverse-kinematics solvers and motion planners that translate model intent into safe, physical joint trajectories.
kinematics.ik → stableOur current build routes fine-tuned Llama models directly into industrial PLCs, translating natural-language maintenance instructions into ladder-logic-safe commands for AC/DC drive controllers — no middleware SCADA layer required.
Added safety interlock layer between agent output and drive controller commands.
Reduced inference latency on edge quantized model from 40ms → 22ms.
Merged ladder-logic translator for Allen-Bradley PLC dialect.
Opened RFC for open kinematics dataset contribution guidelines.
Clone the stack, fork a model, or just come argue with us about kinematics.