Research

Zhang, Y., Hu, Y., Song, Y., Zou, D. & Lin, W. Nature Mach. Intell. (2025).

The primary source that triggered Q2's research direction. Demonstrates autonomous multi-agent flight at 20 m/s with a 90% success rate using differentiable physics simulation, no state estimation, and sub-$21 edge compute. Directly establishes the viability of physics-informed policy learning for the hardware class Q2 targets.

Gao, F. et al. IEEE Trans. Robot. 36, 1526–1545 (2020).

Establishes the complete autonomous flight stack (global planning, local re-planning, and VIO localization) that defines the infrastructure problem Q2's work is designed to operate without.

Zhou, X., Wang, Z., Ye, H., Xu, C. & Gao, F. IEEE Robot. Autom. Lett. 6, 478–485 (2020).

Demonstrates gradient-based local planning without expensive distance field computation, directly informing the efficiency constraints of Q2's edge-compute deployment target.

Zhou, X., Zhu, J., Zhou, H., Xu, C. & Gao, F. arXiv:2011.04183 (2020).

First systematic solution for fully autonomous decentralized quadrotor swarms in unknown cluttered environments using only onboard resources. Extends EGO-Planner with implicit topological planning, reciprocal collision avoidance as a soft barrier penalty, and VIO drift correction via agent detection in depth images.

Loquercio, A. et al. Sci. Robot. 6, eabg5810 (2021).

Proves that lean neural networks operating on minimal sensor input can support competent high-speed navigation, motivating Q2's efficiency-first policy architecture.

Zhang, Z. & Scaramuzza, D. ICRA 2534–2541 (IEEE, 2018).

Integrates perception quality directly into trajectory optimization, a complementary approach to Q2's anomaly-observation confidence model.

Foehn, P. et al. arXiv:2005.12813 (2020).

Documents the DARPA-backed AlphaPilot Challenge system, establishing aggressive vision-based flight benchmarks at speeds exceeding 8 m/s. Provides the open-access American performance baseline that contextualizes the ceiling Q2's GPS-denied framework is designed to operate within.

Wang, W. et al. IROS 4909–4916 (IEEE, 2020).

Diverse simulation dataset spanning degraded visual conditions, validating the simulation-first approach Q2 uses for zero-shot physical generalization.

Teed, Z. & Deng, J. NeurIPS (2021).

Establishes the mature visual SLAM baseline for onboard localization across camera modalities without external reference. Q2's anomaly-based approach deliberately departs from this to remove the camera dependency entirely.

Cioffi, G., Bauersfeld, L., Kaufmann, E. & Scaramuzza, D. IEEE Robot. Autom. Lett. 8, 2684–2691 (2023).

Inertial-only odometry for drone racing validates dead-reckoning under GPS denial, directly informing Q2's sensor-minimal navigation assumption.

Kaufmann, E. et al. Nature 620, 982–987 (2023).

Sets the upper bound of learned flight policy capability at human-champion level, demonstrating that PPO-trained policies with empirical sim-to-real residual models can achieve world-champion performance on embedded hardware.

Kaufmann, E. et al. Robotics: Science and Systems (2020).

Demonstrates extreme maneuver learning transferable from simulation, establishing the sim-to-real transfer methodology Q2 applies to swarm coordination policies.

Foehn, P. et al. Sci. Robot. 7, eabl6259 (2022).

Open-source agile quadrotor platform that provides the physical deployment reference architecture and performance envelope for Q2's hardware target.

Liang, J. & Lin, M. C. ICLR Workshop (2020).

Physics-based simulation for differentiable policy training, the foundational technique behind Q2's digital validation methodology and zero-shot generalization approach.

Schulman, J. et al. arXiv:1707.06347 (2017).

The proximal policy optimization algorithm Q2 applies for reinforcement learning-based policy training within the swarm coordination framework.

Giusti, A. et al. IEEE Robot. Autom. Lett. 1, 661–667 (2015).

Early demonstration of visual navigation policy learning in unstructured environments with no pre-labeled map, a direct analogue to Q2's object-blind positional confidence approach.

Shi, B., Bai, X. & Yao, C. IEEE Trans. Pattern Anal. Mach. Intell. 39, 2298–2304 (2016).

End-to-end sequence recognition architecture that informs the temporal modeling of successive anomaly observations in Q2's Markov chain confidence framework.

Cho, K. et al. arXiv:1406.1078 (2014).

Foundational sequence-to-sequence architecture underlying the recurrent modeling of state transition probabilities in Q2's confidence update model.

Qin, T., Li, P. & Shen, S. IEEE Trans. Robot. 34, 1004–1020 (2018).

The most widely deployed GPS-denied localization baseline in the field. Informs Q2's multi-sensor integration approach: low-resolution camera frames evaluating terrain conformation with computationally simple algorithms improve positional accuracy and precision beyond inertial sensing alone.

Paszke, A. et al. NeurIPS (2019).

The deep learning framework used for implementing and training Q2's policy networks and confidence estimation models.

Loquercio, A. et al. IEEE Trans. Robot. 36, 1–14 (2019).

Domain randomization for sim-to-real transfer in aggressive flight, directly informing Q2's zero-shot generalization methodology from digital simulation to physical deployment.

Metz, L. et al. arXiv:2111.05803 (2021).

Characterizes failure modes of gradient-based optimization in chaotic dynamical systems, informing Q2's hybrid approach to policy training under non-smooth swarm dynamics.

Bengio, Y., Simard, P. & Frasconi, P. IEEE Trans. Neural Netw. 5, 157–166 (1994).

Foundational characterization of the vanishing gradient problem in recurrent networks, directly relevant to Q2's Markov chain formulation for iterative confidence updating across distributed nodes.

Groves, K., Hernandez, E., West, A., Wright, T. & Lennox, B. Robotics 10, 78 (2021).

Converts radiation measurements into a navigation cost signal so a ground robot can actively avoid high-dose areas while exploring an unmapped facility. Direct analogue to Q2's anomaly-observation confidence model, with dose rate replacing terrain anomaly as the observed variable.

West, A., Wright, T., Tsitsimpelis, I., Groves, K., Joyce, M. J. & Lennox, B. Front. Robot. AI 9, 862067 (2022).

Layered costmap formulation for real-time radiation avoidance during exploration of poorly documented environments, informing how Q2's confidence-based path planning would need to weight an actively updated hazard field rather than a static preloaded prior.

West, A., Tsitsimpelis, I., Licata, M., Jazbec, A., Snoj, L., Joyce, M. J. & Lennox, B. Sci. Rep. 11, 13975 (2021).

Interpolates sparse, noisy radiation measurements into a continuous reactor dose-rate map, benchmarked against MCNP6 transport simulation. Establishes the sensor-fusion baseline Q2's radiation-informed confidence model would need to match or improve on.

Stibinger, P., Baca, T., Doubravova, D., Rusnak, J., Solc, J., Jakubek, J., Stepan, P. & Saska, M. arXiv:2510.26018 (2025).

Cooperative multi-drone radiation-source localization using miniature Compton cameras, bringing Q2's existing swarm-coordination and distributed-confidence background directly into the radiation-sensing domain.

Będkowski, J. SoftwareX (2024).

Open-source, feature-less LiDAR/IMU SLAM stack purpose-built for nuclear facility inspection, the closest existing analogue to the indoor, GPS-denied localization problem Q2's terrain-prior framework was designed around.

Vairagade, H., Kim, S., Son, H. & Zhang, F. Front. Energy Res. 12, 1356624 (2024).

Full-scope NPP simulator driving a 3D digital twin for training inspection robot swarms. Validates the simulation-first, zero-shot-to-physical methodology Q2 already applies to its swarm navigation work, applied instead to reactor interiors.

Zhang, K., Hutson, C., Knighton, J., Herrmann, G. & Scott, T. Front. Robot. AI 7, 6 (2020).

Controlled gamma exposure testing of a commercial manipulator, establishing a methodology for characterizing edge-compute and actuator survivability limits before physical deployment in an activated environment.

Nancekievill, M. et al. Sensors 19, 4602 (2019).

Field-validated localization system for an underwater inspection robot at a full-scale Fukushima Daiichi test mockup, grounding the reactor-robotics direction in a real deployed accident-response precedent rather than a lab demonstration only.

Damiani, C. et al. Fusion Eng. Des. (2018).

Defines the remote-handling architecture and 30-year operational lifespan constraints that ITER-class fusion facilities impose on maintenance robotics, establishing the scale and durability requirements this research direction has to eventually address.

Yin, R., Wu, H., Li, M., Cheng, Y., Song, Y. & Handroos, H. arXiv:2208.11737 (2022).

Learned autonomous assembly for plasma-facing component replacement, representing the manipulation half of fusion maintenance robotics that complements Q2's navigation-focused contribution.

Jonasson, E. T., Pinto, L. R. & Vale, A. Fusion Eng. Des. 172, 112691 (2021).

Benchmarks depth camera, LIDAR, and mmWave RADAR for localization inside nuclear facilities, informing which sensing modalities remain viable when radiation and confined geometry degrade the options Q2's drone-domain work assumed were available.

Benito, T. & Barrientos, A. Appl. Sci. 14, 4814 (2024).

Long-lifecycle HMI architecture (MAMIC) for coordinating evolving remote-handling equipment over a multi-decade fusion facility lifespan, framing the systems-integration problem any autonomous navigation layer has to plug into.

Patel, H. K. & Rathod, P. J. Int. J. Sci. Res. 11, 1524–1527 (2022).

Surveys robotic and automation applications across SMR/micro-reactor manufacturing, inspection, and maintenance, the clearest existing statement of the SMR-specific opportunity this research direction targets.

Jog, N. M.S. thesis, Carnegie Mellon University Robotics Institute (2019).

Demonstrates a robot (NanoPiper) built to operate inside 3-inch nuclear piping, using acoustic localization and miniaturized radiation sensing. Sets the miniaturization floor for edge-compute hardware operating in the tightest reactor-interior geometries.