Telekinesis.ai

Telekinesis GmbH, based in Darmstadt, Germany, develops what it calls an Agentic OS for Physical AI: a full-stack software platform and Python SDK intended to let engineers build, simulate and deploy applications across robotic arms, humanoids and mobile bases without rewriting code for each platform. The stack is distributed as a pip-installable package (telekinesis-ai) and bundles more than 250 models and skills spanning perception, planning, manipulation, control and learning, aimed at manufacturing and logistics use cases.

The architecture has four stated layers: an agent built on VLM/LLM models that uses a code-as-policy approach to translate natural-language prompts into auditable Python robot programs; a modular skill library covering perception, motion and planning; a data engine for structuring and generating training data, including synthetic photorealistic labeled datasets produced with domain randomization; and a hardware abstraction layer built around an in-house low-latency robotics communication framework called BabyROS for cloud-to-robot connectivity. Applications marketed on the company's site include bin picking, pick and place, palletizing and depalletizing, assembly, machine tending, sorting, packaging, material handling and label classification.

A company profile on the German business network XING describes the company as developing an AI technology termed Visual Robot Programming, which is intended to let robots learn behaviors autonomously by watching videos, with a stated aim of enabling non-specialists to teach robots through a no-code approach.