Robotics companies in DACH often assume robot software cannot be built remotely because the robot sits in Munich or Zurich. In practice, a large part of a modern ROS 2 stack is ordinary software: nodes, message interfaces, planners, perception pipelines, launch configuration, tests and tooling. These can be built and verified away from the hardware, provided the simulation and replay infrastructure is good enough. The work that genuinely needs the robot is narrower than most teams expect, and it can be scheduled. This note sets out how to draw the line, and what infrastructure must exist before a remote team is productive.
The robotics context in numbers
Demand for robot software is growing faster than the teams that write it. The IFR World Robotics 2025 report puts Germany's robot density at 449 industrial robots per 10,000 manufacturing employees, fourth worldwide, and counts 26,982 new installations, the highest in Europe. The next wave is already visible: in Bitkom's 2026 Industrie 4.0 study, 6 percent of German manufacturers with 100 or more employees already work with humanoid robots and a further 10 percent plan to, while 40 percent use AI in production. Each of these systems adds perception, planning and integration code. Engineering supply has not kept pace. Bitkom counted 79,000 unfilled IT positions in Germany in 2026, while NASSCOM estimates that India's technology industry employed 5.8 million people in fiscal year 2025. Remote ROS 2 work is one way to close that gap, but only if the split between simulation and hardware is designed deliberately.
Start with the stack, not the org chart
A typical ROS 2 system for a mobile manipulator or autonomous mobile robot has six layers. Each has a different dependency on physical hardware.
| Layer | Typical components | Remote suitability |
|---|---|---|
| Mission and fleet logic | Behaviour trees, task allocation, fleet interfaces | High. Pure software, testable in simulation |
| Planning | Nav2, MoveIt 2, custom planners | High. Simulation and recorded maps cover most cases |
| Perception | Detection, segmentation, localisation, SLAM | High to medium. Depends on recorded sensor data quality |
| Control | ros2_control controllers, trajectory execution | Medium. Tuning needs real dynamics |
| Hardware interfaces | ros2_control hardware components, drivers, fieldbus | Low to medium. Needs a rig or the robot |
| Safety functions | Safety PLC, safety-rated stops, speed and separation monitoring | Low. Belongs to the manufacturer's safety process |
The upper three layers are where a remote team creates most value. The lower two need a hardware path. The safety layer should be kept architecturally separate from the ROS 2 stack and handled under its own process, which we cover in a separate article on the Machinery Regulation.
The infrastructure that makes remote ROS 2 work
A remote team is only as productive as its ability to run the system without the robot. Four pieces are essential.
- A maintained simulation. Gazebo (the current generation, since Gazebo Classic reached end of life in January 2025) or NVIDIA Isaac Sim, with robot description files that are kept in sync with the real hardware. If the URDF drifts from the physical robot, simulation results stop meaning anything.
- Headless simulation in CI. Every merge request runs scenario tests in simulation: navigate to goal, pick from a known pose, recover from a blocked path. Results are pass or fail against defined tolerances.
- A rosbag2 library. Recorded sensor data from real runs, tagged by scenario and environment, used for perception regression tests and for reproducing field issues. This is the single most valuable asset for remote perception work.
- A remotely accessible hardware-in-the-loop rig. Not the full robot, but the compute unit, sensors and actuators that matter, mounted on a bench in your lab with network access, power control and a camera. Access is scheduled and logged.
Middleware across networks
Running DDS discovery across a VPN between Bangalore and a lab in Germany is a frequent source of wasted days. Multicast discovery does not traverse most VPNs, and default QoS settings behave differently over high-latency links. Decide early whether the remote team connects to the rig through a discovery server, a bridge, or by running the ROS graph entirely on the rig and accessing it through a remote session. Recent ROS 2 distributions also offer rmw_zenoh as an alternative middleware that handles routed networks more naturally. Whatever you choose, document it as part of the developer setup and test it in the first week.
Distribution and version discipline
Pin the ROS 2 distribution and keep both teams on the same one. Long-term-support releases (currently Jazzy Jalisco from 2024 and Lyrical Luth from May 2026) are the sensible default for products. Use containerised development environments so that a node built in Bangalore and a node built in Stuttgart link against identical packages. Most cross-site integration bugs we see described in the community are version or configuration mismatches, not logic errors.
Acceptance criteria that work remotely
Robotics acceptance criteria are often vague (the robot should navigate reliably). Under a Werkvertrag with sprint-based acceptance, they need to be measurable and checkable without a debate. Examples:
- Navigation: 50 consecutive simulated runs across the five reference maps with zero collisions and a goal tolerance of 0.10 m.
- Perception: recall of at least the agreed threshold on the tagged rosbag2 regression set, with no regression of more than one percentage point on any subset.
- Recovery: from each of the eight defined blocked-path scenarios, the robot reaches a safe state within the agreed time.
- Hardware path: the new driver passes the rig test sequence, and a 30-minute soak test on the rig produces no dropped messages above the agreed rate.
The numbers above are illustrative. The point is that each criterion names the environment (simulation, replay, rig, robot), the dataset, and the threshold.
A step sequence for the first six weeks
- Week 1: containerised dev environment, simulation running headless, CI skeleton, access to the rosbag2 library.
- Week 2: first scenario tests in CI. Agree the reference maps and scenarios with your robotics lead.
- Week 3: HIL rig online with remote access, power cycling and camera. First driver or interface task on the rig.
- Week 4: first feature delivered against simulation and replay criteria.
- Week 5: on-site validation window on the real robot, run by your team with the remote team connected.
- Week 6: review which failures were caught in simulation and which only appeared on the robot. Improve the simulation where it missed something.
Three failure modes to avoid
- Simulation as an afterthought. If nobody owns the simulation, it decays within weeks and the remote team falls back to waiting for hardware time. Assign ownership explicitly.
- Unlabelled recordings. A folder of rosbag2 files without scenario tags, calibration versions and hardware revisions is close to useless for regression testing.
- Rig access without a calendar. When the HIL rig is shared informally between local and remote engineers, the remote side loses. Book it like a meeting room.
What stays with you
Physical testing on the full robot, safety validation, field deployment and customer site acceptance remain on site. The remote team prepares, instruments and supports these steps.
Remote robotics engineering works when the hardware windows are planned and the simulation is trusted. A niogod™ pod is structured for this split: the Projektleiter coordinates on-site test windows in German with your lab, while the engineers deliver against simulation and the rig under a Werkvertrag with Abnahme per sprint, as described on the pod page. If a new robot platform or a ROS 2 migration is planned for next year, it is a conversation worth having early.
Talk to niogod™
Planning engineering capacity in DACH? Book a 30-minute consultation with the niogod™ team.
Book consultation