In plain words
The researchers trained a robot navigation policy in simulation and tested it on a physical iw.hub. With facility-specific data and Cosmos visual variants, 91.7% of trials met the reported arrival condition, compared with 65.3% for training only in NVIDIA environments.
That condition requires reaching within 0.5 m of the goal at linear speed below 0.1 m/s and angular speed below 0.1 rad/s, without collision or timeout.
The research
Can a robot navigation system trained in simulation still work reliably in a real industrial facility, where images, physics and sensor noise differ? The paper tests this question on an Idealworks robot, including whether data from a facility copy and Cosmos-generated visual variations help.
The paper extends COMPASS to Idealworks robots and tests whether facility-specific simulation data, with Cosmos visual variants, improves real-world navigation.
The study tests how simulation training transfers to navigation on a physical iw. hub robot. Across four scenarios, facility-specific data and Cosmos visual variants achieved 91.7% success, against 65.3% for NVIDIA-environment-only training.
Success means reaching within 0.5 m of the goal at linear speed below 0.1 m/s and angular speed below 0.1 rad/s, without collision or timeout. These results compare the tested navigation policies and scenario setup; they do not establish general safety or performance on every industrial route.
The physical tests covered one robot embodiment in one R&D center, and the comparison between training versions changes the resulting base and residual policies as well as the data. The results do not establish performance across other robots or facilities.
Analysis
Deep
- Original paper authors
- Charbel Abi Hana, Rayan Zaitouni, Kameel Amareen, Mohamad Mostafa, Hasan Sayour
- Submitted to arXiv
- License
- CC BY 4.0