SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Rohlik is Central Europe's leading e-grocer, operating across five countries with over a million customers. The company is building autonomous fulfillment centers using AI, vision models, and robotics to deliver fresh groceries at unprecedented cost efficiency.
You will lead the machine-learning side of a newly created robot-learning team focused on capturing and labeling human manipulation data in warehouses, then training robot policies on that corpus. This is a hands-on, high-impact role where your technical judgment directly influences capital allocation and product direction.
Key responsibilities:
- Own the capture specification: define camera placement, calibration standards, and episode-quality criteria before hardware procurement. Validate the spec against real warehouse conditions and iterate based on floor reality.
- Establish ground-truth labeling for hand pose and manipulation, accounting for production challenges (work gloves, occlusion, cold environments) where published models typically fail.
- Build the evaluation harness that determines which recorded hours are trainable and meet quality thresholds as volume scales.
- Execute first training runs: post-train open-source robot-learning models on Rohlik's proprietary warehouse data and run initial task evaluations.
- Stay current with the field: read and reproduce published claims, validate assumptions before budget is spent, and steer capture strategy based on emerging research.
You will work on a small team: an operations lead managing floor capture, a data engineer owning the pipeline, and you as the ML voice making final calls on data quality and trainability. This is not a research role; the focus is building a production-grade training corpus and demonstrating policy learning on real warehouse tasks.
The role emphasizes pragmatism: you will use AI agents (Claude, Devin) for routine engineering work while maintaining rigorous standards for the critical path. Success means writing a capture spec, setting up initial rigs, establishing pilot gates with measurable criteria, and providing data-backed recommendations for scaling to hundreds of hours of training data.
REQUIREMENTS:
- Strong PyTorch and computer vision fundamentals, including 3D geometry, camera calibration, pose estimation, and SLAM basics. Hands-on experience debugging extrinsic calibration drift.
- Ability to read academic papers and reproduce claims in real-world conditions.
- Familiarity with current robot-learning stacks and policy classes; understanding of what training pipelines demand from data before collection begins.
- Direct experience training on self-collected datasets and knowledge of failure modes that emerge at training time.
- Comfort operating as the sole ML decision-maker on a small team, documenting decisions and revising based on pilot results.
- Proficiency with AI-assisted engineering (Claude Code, Devin) for non-critical tasks.
- Particularly valued: hand-pose estimation or egocentric video in production systems; multi-camera rig design and time synchronization; experience post-training robot-learning models; ROS 2 familiarity.
