> Markdown version of [/jobs/ext/4121342-machine-learning-engineer-vla-training-evaluation](https://www.wearedevelopers.com/jobs/ext/4121342-machine-learning-engineer-vla-training-evaluation). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# Machine Learning Engineer VLA Training & Evaluation

- **Company:** Agile Robots Ag
- **Location:** München, Germany
- **Contract:** Permanent contract
- **Skills:** Agile Methodology, Artificial Intelligence, Extract Transform Load (ETL), Software Debugging, Machine Learning, Language Modeling, Performance Tuning, Management of Software Versions, Pytorch, Deep Learning, Information Technology, Software Version Control
- **Published:** October 15, 2026
- **Apply:** https://www.indeed.com/viewjob?jk=17898b7603646717

## About the Role

* Academic Background: Degree in Computer Science, Machine Learning, Robotics, or a related field (Master's or PhD typical), or an equivalent combination of formal training and demonstrated experience training deep learning models.
 * Training Internals: Hands-on experience writing and debugging training loops in PyTorch or JAX, including data loading, loss design, and optimizer behavior.
 * Fine-Tuning: Experience fine-tuning pretrained models and adapting them to new tasks and data.
 * Robot Learning: Working understanding of how imitation-learned policies, such as action-chunking or diffusion policies, turn observations into robot actions.
 * Experiment Management: Experience designing evaluations and ablations, and running experiments with tracking, versioning, and checkpoint management so that others can reproduce them., * VLA Models: Experience fine-tuning vision-language-action or vision-language models for robot control.
 * Real-Robot Evaluation: Experience evaluating learned policies on physical robots.
 * Sim-to-Real: Familiarity with measuring how well models trained in simulation transfer to real hardware.

## Description

* Training Workflows: Build reusable, model-agnostic training and evaluation workflows, with experiment tracking, model versioning, and checkpoint management, so that different foundation models can be swapped in without rebuilding the pipeline.
 * Fine-Tuning: Implement fine-tuning of pretrained vision-language-action models on collected robot data, for example with parameter-efficient methods such as LoRA or adapters, or with distillation.
 * Policy Training: Build training loops for action-chunking and transformer-based action models, diffusion policies, and VLM/VLA heads.
 * Benchmarks: Build benchmarks with generalization splits, success metrics, and failure-analysis templates.
 * Transfer Studies: Work with simulation engineers on studies, including metrics and ablations, that test whether improvements in simulation realism improve performance on the real robot.
 * Research Baselines: Turn published methods into working baselines and templates that others can extend.