Software stack for Physical AI: training, simulation and deployment of general-purpose industrial robots via a unified Agentic OS

Telekinesis GmbH

Darmstadt (Headquarters) · Frankfurt am Main (Team/Office)New on Amplifa

GmbH, HRB 105059 (Darmstadt Local Court)Legal form/Register

Suman Pal, Arjun Vir DattaManagement

250+Skills in the library

10+supported hardware OEMs

About Telekinesis

Telekinesis GmbH is a software company from Darmstadt that develops a standardized operating system platform (Agentic OS) for Physical AI. The offering includes a Python SDK (telekinesis-ai) for training, simulation and deployment of robots, a Code-as-Policy agent that translates natural language prompts into executable, auditable Python code, and a library with over 250 skills for perception, planning, manipulation, control and reinforcement learning. Applications include bin picking, pick & place, palletizing/depalletizing, sorting, packaging, label classification, assembly, machine tending and material handling. The software supports robot arms, mobile robots and humanoids from various manufacturers via a Hardware Abstraction Layer (BabyROS); simulation is performed with NVIDIA Isaac Sim, Isaac Lab and MuJoCo. Telekinesis is the first spin-off of the Intelligent Autonomous Systems Lab at TU Darmstadt (Prof. Jan Peters); the team is based in Frankfurt am Main and is backed by Atlantic Labs as well as business angels, including former CEOs of KUKA and Universal Robots. Target industries are robotics manufacturers, manufacturing and logistics.

Site: Darmstadt
Processes: Physical AI SDK (telekinesis-ai) for training, simulation and deployment of robots, Agentic OS / Code-as-Policy agent (prompt to executable Python code), Skill library (250+ skills): perception, planning, manipulation, control, reinforcement learning, 3D point cloud processing, image segmentation, object detection, image processing, Synthetic data generation (extreme domain randomization), Hardware Abstraction Layer / robotics messaging (BabyROS), Simulation with NVIDIA Isaac Sim, Isaac Lab, MuJoCo, Open-source components: BabyROS, RLBotics (Apache 2.0), Illusion (GPLv3)
Materials: Physical AI SDK (telekinesis-ai), Agentic OS, Code-as-Policy Agent, Skill Library (250+ skills), Simulation (Isaac Sim, Isaac Lab, MuJoCo), Hardware Abstraction Layer (BabyROS), Synthetic Data Generation, Open-source components
Certificates: Open-source licenses (Apache 2.0, GPLv3), ISO 9001
Industries: Robotics / Robotics manufacturers, Manufacturing / Production, Logistics

01
Prompt/Natural language
02
Code-as-Policy agent
03
Skill library (perception, planning, manipulation)
04
Simulation (NVIDIA Isaac Sim, Isaac Lab, MuJoCo)
05
Deployment on any robotics hardware (BabyROS)
06
Fleet operation in production

Software, Skills & Processes

SDK

Physical AI SDK (telekinesis-ai)

Python SDK for training, simulation and deployment of general-purpose robots via a pip install command.

Agentic OS

Complete software stack that connects robotics manufacturers and production cells via unified interfaces.

VLM/LLM

Code-as-Policy Agent

Translates natural language prompts into readable, auditable Python code for robot motions.

SKILLS

Skill Library (250+ skills)

Modular capabilities for perception, planning, manipulation, control and reinforcement learning, including 3D point cloud processing, image segmentation and object detection.

SIM

Simulation (Isaac Sim, Isaac Lab, MuJoCo)

Validation of generated applications in simulation before deployment on real hardware.

HAL

Hardware Abstraction Layer (BabyROS)

Ultra-low-latency pub/sub and client/server messaging from cloud to robot across OEM boundaries.

DATA

Synthetic Data Generation

Photorealistic, perfectly labeled datasets using extreme domain randomization.

OSS

Open-source components

BabyROS and RLBotics (Apache 2.0) as well as Illusion (GPLv3) are publicly released as open source.

Locations

Darmstadt

Headquarters / Office (Neckarstraße 4, Room 4.1.08)

Materials: Agentic OS, SDK, skill library

Cloud platform
Simulations (NVIDIA Isaac Sim, Isaac Lab, MuJoCo)

Frankfurt am Main

Team / Office (WeWork, Taunusanlage 8)

Materials: Software development

Applications & Architecture

**Bin picking with 6-axis collaborative arm**Physical AI SDK · Manufacturing

**Sorting & material handling with AMR**Agentic OS · Logistics

**Palletizing with humanoid robot**Skill library · Production

Inside the production

Telekinesis – Agentic OS for Physical AI

Industries

Robotics / Robotics manufacturersManufacturing / ProductionLogistics

Additional services

Physical AI SDK (telekinesis-ai) for training, simulation and deployment of robotsAgentic OS / Code-as-Policy agent (prompt to executable Python code)Skill library (250+ skills): perception, planning, manipulation, control, reinforcement learning3D point cloud processing, image segmentation, object detection, image processingSynthetic data generation (extreme domain randomization)Hardware Abstraction Layer / robotics messaging (BabyROS)Simulation with NVIDIA Isaac Sim, Isaac Lab, MuJoCoOpen-source components: BabyROS, RLBotics (Apache 2.0), Illusion (GPLv3)

Licenses & Open Standards

The platform validates generated applications in simulations (NVIDIA Isaac Sim, Isaac Lab, MuJoCo) before hardware deployment; applications are deployed versioned, traceable and reversible on every machine. Stated key figures are 99% task accuracy and 0.01 mm motion precision.

Evidence	Telekinesis
Open-source licenses (Apache 2.0, GPLv3)
ISO 9001

History

Telekinesis is founded as the first spin-off of the Intelligent Autonomous Systems Lab at TU Darmstadt (Prof. Jan Peters); the team is based in Frankfurt, backed by Atlantic Labs and business angels, including former CEOs of KUKA and Universal Robots.