RatSLAM

The first RatSLAM paper, presented at ACRA 2003, used separate competitive attractor networks for place and head direction. Visual observations recalibrated movement estimates, enabling one-dimensional mapping and self-calibration. Ambiguous two-dimensional environments exposed difficulties in preserving consistent combinations of position and heading. [R1]

ICRA 2004 introduced pose cells: a joint representation of position and heading that addressed the earlier binding problem. Those experiments used artificial landmarks. Later that year, ACRA experiments demonstrated natural-scene recognition in unmodified indoor environments, including a corridor loop and an office floor. [R2][R3]

A goal-memory component learned travel times between place representations during exploration. Following the resulting temporal gradient guided a Pioneer robot towards designated destinations in an office and laboratory. The ICRA 2005 experiments combined real-time SLAM with immediate goal-directed navigation. [R35]

ACRA 2005 introduced an experience graph combining pose, view and movement information. Graph adjustment produced a spatially coherent map despite discontinuities in the neural representation. IROS 2006 evaluated limited pose-cell resources, goal-directed navigation and adaptation to simple environmental changes. [R4][R5]

Outdoor experiments extended RatSLAM to less controlled visual conditions and robot motion. The work appeared in the Field and Service Robotics research cycle and as a 2006 Springer chapter, preceding the larger suburban mapping experiments. [R6]

RatChat associated RatSLAM experience representations with concepts such as rooms and corridors. A single-layer neural network learned these associations through a teacher–student process and recognised them on revisits. The experiments ran offline on representations obtained during Pioneer robot exploration. [R36]

A 66 km drive through St Lucia, Brisbane, supplied webcam images at 10 Hz over approximately 100 minutes. Visual motion estimation and place recognition produced a coherent semimetric road map with more than 51 loop closures, including loops up to 5 km long. Data were collected in 2007; the journal paper appeared in 2008. [R7][R17]

The 2008 book consolidated the neuroscience motivation, computational architecture and navigation experiments. Published as volume 41 of Springer Tracts in Advanced Robotics, it covers simultaneous localisation, mapping and path planning based on hippocampal models. [R8]

Over two weeks, an office robot completed 1,143 delivery tasks to 11 destinations, with one delivery failure and subsequent recovery. It travelled more than 40 km during 37 active hours and recharged autonomously 23 times. SLAM operated alongside navigation and task-selection modules. The study appeared online in 2009 and in the 2010 IJRR volume. [R9]

FAB-MAP + RatSLAM combined probabilistic appearance recognition with pose filtering and experience mapping. The ICRA 2010 study evaluated observations collected at different times of day, integrating an independently developed recognition method with the RatSLAM architecture. [R10]

RatSLAM maintained competing pose hypotheses in ambiguous environments until sequences of observations distinguished them. The 2010 computational neuroscience study proposed that conjunctive grid-like cells could filter uncertain sensory information. This was a computational hypothesis evaluated through modelling and robotic navigation. [R11]

ICRA 2011 experiments used imagery from a micro aerial vehicle. Visual odometry and place recognition from approximately 1,000-pixel images were combined with anticipated image sequences to improve recall and precision. Outdoor observations included low-quality and intermittent visual input. [R12]

A 2011 study integrated multicamera visual odometry with RatSLAM on a 1,300 m dataset. Altitude was added to experience-map nodes and altitude changes to links, while pose cells retained the x, y and heading representation. This extended the map to 2.5D without increasing the pose network’s dimensionality. [R28]

SeqSLAM uses sequences of images for route recognition under substantial appearance change. It is a distinct method from RatSLAM and does not include pose-cell dynamics or experience mapping. [R34][R4]

The 2013 open implementation separated local-view processing, pose cells and experience mapping into ROS components, with a separate visual-odometry node. The publication documented interfaces, parameters and tuning, and provided demonstrations on three public datasets. [R13]

Movement-based calibration estimated self-motion and place-recognition parameters, while online weighting adapted sensor contributions. Tests used several robot and sensor configurations in offices, a university campus and a forest, producing topologically correct maps in most tested conditions. The study appeared online in 2014 and in the 2015 journal volume. [R14]

NeuroSLAM combines three-dimensional grid-cell models and layered head-direction representations in a four-degree-of-freedom architecture. Its multilayered experience map is adjusted after loop closure. Evaluation used synthetic and real indoor and outdoor datasets. Lead researcher Fangwen Yu visited Milford’s lab during this research. [R16]