Develops learning robotics and Physical AI for mid-sized manufacturing – from simulation to edge inference.
eigenmind GmbH
10115 BerlinNew on Amplifa
About n8shyft
eigenmind GmbH operates under the brand n8shyft as a Berlin-based company in the field of Physical AI and learning robotics for the manufacturing industry. The founding team around Anne Decker (CEO) and Dr. Simon Müller develops vision policies that learn contact-rich handling tasks end-to-end from demonstrations – without classical trajectory planning or per-product-variant programming. The company uses simulation as a data factory: digital twins in MuJoCo and NVIDIA Isaac Sim generate training data with domain randomization before real hardware is deployed. Inference is intended to run locally on industrial hardware (Jetson Orin, projected), so no plant data leaves the premises. Currently, a vision ACT policy (17.1 million parameters, over 1,800 demonstrations) solves a loading task end-to-end in simulation; humanoid integration using the Unitree G1 is in preparation. For mid-sized manufacturing operations, the company offers the consulting format 'Die Nachtschicht' (The Night Shift), providing task screening and a 6-month roadmap directly in ongoing production.
- Site
- Berlin
- Processes
- Learning robotics / Physical AI: vision policies for manipulation, Simulation as a data factory (MuJoCo, Isaac Sim, domain randomization), Edge inference on industrial hardware (Jetson Orin, projected), Humanoid integration (Unitree G1) in preparation, Consulting format 'Die Nachtschicht' (The Night Shift): task screening and 6-month roadmap for
- Materials
- Vision-ACT-Policy, Simulation with MuJoCo, NVIDIA Isaac Sim, Domain Randomization, Edge Inference
- Industries
- Mid-sized manufacturing industry
- 01
Task screening in manufacturing
- 02
Data collection from demonstrations
- 03
Simulation & domain randomization
- 04
Policy training
- 05
Edge inference on industrial hardware
- 06
Piloting on real hardware
Processes & Platforms
Vision-ACT-Policy
End-to-end policy (17.1 million parameters) that learns from demonstrations: camera image in, motion out.
Simulation with MuJoCo
Physics simulation for generating training data prior to real hardware deployment.
NVIDIA Isaac Sim
Simulation platform for digital twins and humanoid integration (Unitree G1 in preparation).
Domain Randomization
Randomization of 25+ parameters makes policies robust to the disorder of real manufacturing environments.
Edge Inference
Projected local inference on industrial hardware; p99 latency 21.4 ms, 7/7 readiness tests passed.
Site
Berlin
DEHeadquarters and development (c/o Factory Works GmbH, Rheinsberger Str. 76/77)
- Materials
- Vision-ACT, MuJoCo, Isaac Sim
- Simulation (MuJoCo, Isaac Sim)
- Camera pipeline (86 Hz)
Inside the production
Industries
Additional services
Roadmap & Milestones
- 2026 Q1
Vision-ACT policy solves a loading task end-to-end in simulation (trained from 1,800+ demonstrations).
- 2026 Q1
Edge readiness: all 7/7 tests passed – real-time-capable inference on industrial hardware, local without cloud.
- 2026 Q2
Humanoid integration in preparation: Unitree G1 as simulation platform in Isaac Sim.