Quadruped Robots as an Integrative Research Platform: A Cross-Cutting Survey of Multimodal Perception, World Modelling, Learned Control, Foundation Models, and Sim-to-Real Transfer

Abstract

Quadruped robots have undergone a radical transformation over the past five years, transitioning from academic demonstrators to commercial products deployed in real industrial environments. The enabling factor of this transformation has been the convergence of five research streams: deep learning for motor control, neural environment modelling, foundation model integration for reasoning, GPU-accelerated simulation, and advanced multimodal perception. This survey provides a systematic cross-cutting review of the state-of-the-art in quadruped robotics, organised across these five research dimensions and covering the period 2019–2025. We first survey the principal hardware platforms — Boston Dynamics Spot, ANYbotics ANYmal, Unitree Go1/Go2/B2, MIT Cheetah, and open-source alternatives — analysing their mechanical, sensory, and software characteristics as research platforms. We then examine, for each research dimension: (i) multimodal perception on quadrupeds, including terrain-aware vision, LiDAR-camera fusion, distributed tactile sensing, and fault-tolerant multi-sensor fusion in hostile environments; (ii) SLAM and world modelling for legged robots, with specific attention to locomotion-induced odometry degradation, legged-robot-specific factor graph optimisation (VILENS), and semantic open-vocabulary mapping (ConceptFusion, HOV-SG); (iii) reinforcement learning for locomotion, from the foundational results of Hwangbo et al. (2019) through agile parkour (ANYmal Parkour, Extreme Parkour) and whole-body loco-manipulation to Rapid Motor Adaptation; (iv) foundation model integration — LLM-guided navigation (ViNT, LEGO-Nav), VLA systems for quadrupeds with manipulators (MOMA-Force, GR00T N1), and dual-system architectures (LEGS, NavGPT); and (v) simulation and sim-to-real transfer, including Legged Gym/Isaac Lab, domain randomisation for terrain and dynamics, and Real-to-Sim-to-Real with Gaussian Splatting (GaussGym). A dedicated section documents 23 real-world experimental deployments across six application domains: industrial inspection (ANYmal D at Aker BP, 200,000+ cumulative operational hours), hazardous environments (DARPA SubT Challenge, Sellafield nuclear decommissioning), military and security applications, medical and assistive robotics, precision agriculture, and space exploration (NASA JPL, ESA LEAP). Seven open research challenges are identified: integrated loco-manipulation, navigation in dynamic human environments, continual adaptation, safety certification for neural policies, energy efficiency, human-robot interaction, and distributed multi-robot intelligence.