Mapless, Stabilizing, and Anti-Stuck Flipper Control for Articulated Tracked Robots

I. INTRODUCTION

• The terrain beneath the chassis is sensor-occluded and must be remembered in an elevation map through wheel or LiDAR odometry, whose drift corrupts the estimate.On loose, deformable terrain, track slip and abrupt chassis motion make odometry least reliable; invalid LiDAR returns can add obstacles that do not exist.Source: §I · p. 1
• Hidden rigid obstacles can remain invisible beneath a compliant canopy until the chassis stalls, while dedicated underbody sensors and foliage-penetrating radar do not resolve the case in practice.The onboard map and ground-truth map comparison on a pallet illustrates perception and localization error, including a reflected invalid LiDAR return.Source: §I · pp. 1–2, Figure 1 · p. 1
• SAS combines roll-and-pitch stabilization with reactive anti-stuck digging, using a thresholded motion cue that is computed over a short window and never accumulated.Its cue needs only to distinguish motion from no motion, not provide an accurate or metric estimate.Source: §I · p. 2
• The evaluation compares SAS with twelve map-based and two manual baselines on a twelve-obstacle simulation course, then compares controllers on a real pallet obstacle.The simulation controllers each complete ten runs, with perception reset at each obstacle; the real comparison also uses ten runs per controller.Source: §I · p. 2

A. Manual control

• Manual control avoids a map because the operator sees the terrain from a third-person view, but requires hand control of all four flippers.Some manual approaches use preset flipper modes instead.Source: A. Manual control · p. 2

B. Reactive and geometry-based control

• Reactive and geometry-based controllers compute flipper commands from terrain models, including elevation maps, known stair geometry, signed-distance fields, or sampled occupancy maps.Methods include geometry-based pose prediction, configuration-space planning, hybrid trajectory optimization, and RRT* planning.Source: B. Reactive and geometry-based control · p. 2

C. Imitation learning

• Imitation-learning approaches map elevation-map terrain classifications or learned states to hand-crafted flipper behaviors; one state machine also overlays roll stabilization and stagnation-triggered anti-stuck control.The cited state machine lowers all four flippers to free the robot when stagnation is detected.Source: C. Imitation learning · p. 2

D. Reinforcement learning

• Reinforcement-learning baselines use map-derived terrain or step geometry, with policies and rewards varying in their treatment of pitch, stability, and getting stuck.C-TRAC uses noisy LiDAR and a contact-aware latent representation, with a reward that stabilizes roll and pitch.Source: D. Reinforcement learning · p. 2

E. Attitude stabilization and mapless posture control

• Prior flipper stabilization usually sits on top of an elevation map, while Rocha et al. regulate roll, pitch, and ground clearance from the IMU and flipper joints.C-TRAC stabilizes roll and pitch from a reward computed on map-based simulator state.Source: E. Attitude stabilization and mapless… · p. 2

F. Evaluating traversal quality

• Traversal quality metrics include body smoothness, attitude, and chassis-to-ground clearance measured from the ground-truth map.The compared-controller table identifies map need, stabilization axes, anti-stuck capability, and code provenance as comparison attributes.Source: F. Evaluating traversal quality · p. 2, Table I · p. 3

III. METHOD

• At each 20 Hz control step, SAS reads IMU roll and pitch, four flipper joint angles, commanded body velocity, and a scalar motion cue, then outputs an angular-velocity command for each flipper.The output combines a roll-and-pitch stabilization term with an anti-stuck term that frees the robot when stuck.Source: §III · p. 3

A. Attitude stabilization

• Roll and pitch errors from zero set points generate proportional-integral commands, distributed as left-right differentials for roll and front-rear differentials for pitch.The integrals are limited by anti-windup saturation.Source: A. Attitude stabilization · p. 3

B. Reactive anti-stuck

• The anti-stuck term digs when the robot is commanded to move but its LiDAR-ICP motion cue falls below the stuck threshold; digging drives all four flippers down to lift the body.The cue is measured over a sliding 0.5 s window; after 8 s of continuous digging without release, the controller suspends digging for 2.5 s and drives the flippers back up.Source: B. Reactive anti-stuck · p. 3

C. Directional-priority blend and limits

• The cap lets anti-stuck commands reinforce or nullify stabilization but never reverse its direction on any flipper.When the commands conflict, α = 0.5 allows the anti-stuck command to weaken stabilization by at most half.Source: C. Directional-priority blend and limits · pp. 3–4
• The blended flipper command is saturated to the drive velocity limit of 1.0 rad·s−1.The cap limits an opposing anti-stuck term relative to the stabilization velocity.Source: C. Directional-priority blend and limits · p. 3
• When stabilization yields completely, the anti-stuck command passes through unchanged; when α = 1, it can nullify stabilization.The anti-stuck command still cannot reverse the stabilization command.Source: C. Directional-priority blend and limits · pp. 3–4

A. Platform and onboard pipeline

• The four-track robot’s four independently actuated flippers are also its only driven tracks, and its ROS 2 perception pipeline produces a robot-centric onboard elevation map.An Ouster LiDAR and IMU feed point-cloud deskewing, LiDAR-ICP odometry, and elevation mapping.Source: A. Platform and onboard pipeline · p. 4
• Simulation uses a twelve-obstacle warehouse course, simulated Ouster and IMU sensors, and ground-truth pose and map from the simulator.The simulated onboard-map condition runs the robot’s ICP and elevation-mapping nodes on simulated ROS 2 topics.Source: A. Platform and onboard pipeline · p. 4
• The real-robot comparison repeats the experiment on one pallet between reset zones, using AprilTags for absolute localization in a handcrafted ground-truth map.Source: A. Platform and onboard pipeline · p. 4
• Autonomous runs command 0.6 m·s−1 in simulation and 0.25 m·s−1 on the real robot; SAS override switches stalled robots to anti-stuck dig mode until movement resumes.The two manual references are driven by an operator who controls the robot and all four flippers by hand.Source: A. Platform and onboard pipeline · p. 4

B. Baselines

• The simulation baseline comparison includes an expert manual controller, preset manual modes, and every map-based controller cataloged in Table I.Where code was released, the authors ran the implementation; otherwise, they reimplemented the controller from its paper.Source: B. Baselines · p. 4
• Geometry-based baselines were adapted for four-track morphology and a passive chassis, including changes to contact cases, stability selection, and obstacle schedules.The adaptations cover GFMP, Yuan’s planner, Oehler and von Stryk’s pose predictor, Xu’s HTO, and Zhang’s planner; Rocha et al.’s mapless posture controller was not evaluated because it is not designed for obstacle traversal.Source: B. Baselines · p. 4
• Learned baselines were adapted for four-track driving, including mirrored rear-flipper rewards, movement incentives, and changes to trailing-flipper rules.The passage describes adaptations to AT-D3QN, ICM-D3QN, C-TRAC, Mitriakov, and Pecka.Source: B. Baselines · p. 4

C. Ground-truth map for a fair baseline

• Each map-based baseline is run with both the onboard map and a ground-truth map, and the score difference measures its sensitivity to onboard mapping.Source: C. Ground-truth map for a fair baseline · p. 4
• The warehouse comparison course contains twelve numbered obstacles of variable difficulty, with a one-metre settling strip between obstacles.Source: Figure 2 · p. 5

D. Metrics

• Traversal quality (TQ) is the mean of three normalized, distance-binned channel scores: shock, body tilt, and ground clearance.The channels represent impacts transmitted to the body, departure from gravity-aligned horizontal, and chassis clearance from the terrain, respectively.Source: D. Metrics · p. 5
• Shock is normalized IMU acceleration magnitude, tilt is the angle between the body z axis and vertical, and ground clearance is measured from a ground-truth map beneath the chassis.Real-robot shock uses a maximum measured acceleration of 57.5 m·s−2; simulated shock uses a maximum 0.1 s mean of 376 m·s−2.Source: D. Metrics · p. 5
• Each channel uses a sigmoid score, and the worst per-sample score in each 0.25 m distance bin is averaged across bins.Distance binning is used because controllers take different times to traverse the same obstacle.Source: D. Metrics · p. 5
• Pairwise correlations among the channel scores are at most 0.39 over traversed obstacles, suggesting that none of the evaluation channels are redundant.Table II reports correlations over simulated obstacles and real crossings.Source: D. Metrics · p. 5, Table II · p. 5

E. Simulation results

• 0.89, manual control's TQ, was 0.05 above the best controller in simulation, with most of the gap in clearance (0.87 versus controllers' 0.61–0.75).Every controller lost 0.008–0.069 under the onboard map; SAS lost 0.020 through its LiDAR-ICP motion cue alone. Most controllers used the override for 7%–15% of obstacle distance under either map.Source: E. Simulation results · p. 5, Table III · p. 6

F. Real-robot results

• 0.81, both manual rides' TQ on the real pallet, exceeded every controller; SAS ranked second under the onboard map, behind Yuan (0.72).Oehler and Mitriakov tied at 0.71 under the ground-truth map.Source: Table III · p. 6, F. Real-robot results · p. 6

G. Hidden obstacle

• On a log hidden in tall grass, Yuan lowered its front flippers and became stuck beneath the log, while SAS crossed it.The onboard map represented the grass canopy and omitted the log.Source: G. Hidden obstacle · p. 6

H. Onboard-map error

• Onboard-map error peaked five times higher on the real robot than in simulation, and cells off by more than 10 cm were fifteen times more common.Error rose during ascent, when terrain beneath the body was last observed from the flat approach, then fell as terrain ahead reappeared during descent.Source: H. Onboard-map error · p. 6

• On the real pallet, clearance is the most significant factor separating controllers, and every controller’s clearance score drops from 0.68–0.75 in simulation to 0.30–0.55 on the real robot.Source: §V · p. 6
• On a real log hidden in tall grass and absent from the onboard map, Yuan gets stuck underneath it while SAS crosses.Source: Figure 3 · p. 7, §V · p. 7
• Every map-based controller scores 0.008 to 0.069 lower under the onboard map than the ground-truth map in simulation, compared with a 0.020 loss for SAS using its motion cue.Source: §V · p. 7
• SAS achieves the highest traversal quality of all autonomous controllers in simulation under both maps, and ranks second under the onboard map and among the leading controllers under the ground-truth map on the real pallet.Source: §V · p. 7
• The comparisons reset the perception stack at each obstacle, making the reported map degradation a lower bound and leaving long, uninterrupted traversal untested.Source: §V · p. 7