Towards collaborative object transportation using differential drive robots : A graph neural network based approach
Guo, Rong (2026)
2026
Master's Programme in Automation Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-07-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202607248461
https://urn.fi/URN:NBN:fi:tuni-202607248461
Tiivistelmä
Collaborative object transportation requires groups of robots to carry objects to their goals while holding each object rigidly and avoiding collisions with other groups. The task becomes difficult as the number of groups grows: the groups need to coordinate in a shared space and keep formation while avoiding others in dense scenes.
A common approach uses an optimization framework, for example a Quadratic Program with Control Barrier Functions (QP-CBF), to enforce the formation and the safety distance as hard constraints. This gives safety guarantees for a small number of groups, but the hard constraints can conflict when many groups share a small space, leaving the solver with no feasible action. The groups then deadlock and fail to reach their goals. The per-step optimization also grows more expensive as the number of groups increases.
This thesis investigates whether a Graph Neural Network (GNN) can take over the task and scale to denser scenes. The system consists of groups of two differential-drive robots, where each group carries one object to its goal while avoiding the other groups. A GNN policy is trained using the QP-CBF controller as an expert, with physics-based loss terms, and is evaluated against the expert in single-group and multi-group scenarios.
In the single-group task, the policy reaches the goal in nearly all runs, the same as the expert, with a small loss of formation precision. In the multi-group task, the expert's success rate falls from 100% with two groups to 4% with eight, while the policy stays above 82% through ten groups, including group counts beyond its training range. The policy avoids the deadlock by turning away from other groups early and keeps working beyond its training range because each group only reacts to its nearby neighbors. It gives up some formation precision but keeps the groups moving and reaches the goals where the expert becomes infeasible.
A common approach uses an optimization framework, for example a Quadratic Program with Control Barrier Functions (QP-CBF), to enforce the formation and the safety distance as hard constraints. This gives safety guarantees for a small number of groups, but the hard constraints can conflict when many groups share a small space, leaving the solver with no feasible action. The groups then deadlock and fail to reach their goals. The per-step optimization also grows more expensive as the number of groups increases.
This thesis investigates whether a Graph Neural Network (GNN) can take over the task and scale to denser scenes. The system consists of groups of two differential-drive robots, where each group carries one object to its goal while avoiding the other groups. A GNN policy is trained using the QP-CBF controller as an expert, with physics-based loss terms, and is evaluated against the expert in single-group and multi-group scenarios.
In the single-group task, the policy reaches the goal in nearly all runs, the same as the expert, with a small loss of formation precision. In the multi-group task, the expert's success rate falls from 100% with two groups to 4% with eight, while the policy stays above 82% through ten groups, including group counts beyond its training range. The policy avoids the deadlock by turning away from other groups early and keeps working beyond its training range because each group only reacts to its nearby neighbors. It gives up some formation precision but keeps the groups moving and reaches the goals where the expert becomes infeasible.