Optimus by Iacon Autonomics

Optimus by Iacon Autonomics is a research tool. Optimus from Iacon Autonomics is an autonomous reinforcement learning engineer. Given an agent or model and a goal, it researches methods, runs experiments and trains LLMs, agents and robot policies, with desktop apps for Mac, Windows and Linux.

Optimus, operated by Iacon Autonomics, is an autonomous reinforcement learning engineer. A team supplies an agent or model together with a goal, and Optimus researches algorithms, designs experiments, runs them and evaluates which changes actually improve the result. The stated scope runs from fine-tuning language models to training robot policies, and the company also names diffusion models, NPCs, biological systems and custom policies as targets. The product is organized as a training stack that Optimus operates end to end. It starts with data, turning raw trajectories into higher-signal training sets. It then rebuilds the customer's real environment as an RL environment where the agent can practise, applies reinforcement learning against rewards, fine-tunes on the most instructive examples, and runs post-training to shift outputs toward preferred answers. Benchmarking compares each checkpoint with a baseline task by task, and evaluation promotes a checkpoint only after every check has passed. Optimus also decides what to try next across those layers. In practice, humans keep control of direction: they define the environments, rewards and success criteria, while Optimus runs the loop inside the customer's own stack. It can implement published state-of-the-art methods or synthesize new ones. Desktop downloads are offered for Mac (Apple Silicon), Windows (x64) and Linux (x64 AppImage). The site presents case studies with Resemble AI (an investigative-agent benchmark called DetectBench), Google DeepMind (teaching Gemma to paint through code) and an independent MuJoCo project for a Franka Panda picking task. Several research groups and companies are listed as users, and a partnership program invites compute, simulator and data providers to integrate with the stack. Among alternatives, Optimus sits closer to an automated research engineer than to a training library or hosted fine-tuning service. It suits teams that already have a defined environment and reward and want the experimentation loop itself delegated, rather than people looking for a one-click model customization tool.