All-Wheel Omni-Directional Independent Steering Vehicle

• AWOISV is a vehicle platform with independently steerable wheels capable of ±90° rotation, enabling zero-radius turning and omnidirectional motion.
• It employs advanced control strategies, including MPC, fuzzy logic, and DRL, to manage redundant actuation and enforce mechanical feasibility.
• Practical applications in autonomous mobile robotics have shown robust path tracking, improved stability, and safe operation in industrial settings.

An all-wheel omni-directional independent steering vehicle (AWOISV) is a specialized all-wheel independent steering vehicle in which each wheel can steer independently over the full range $(-\pi/2,\pi/2)$, i.e., up to $\pm 90^\circ$, enabling zero-radius turning, diagonal motion, lateral motion, yaw-in-place-like motion, and arbitrary combinations of translation and rotation (Yang et al., 19 Aug 2025). In closely related mobile-robot terminology, a four-wheel independent steering and driving (4WISD) autonomous mobile robot is described as a platform in which each wheel is independently steerable and drivable, with omnidirectional motion and zero-radius turns; in practical terms, this is the same vehicle family often meant by AWOISV, although the autonomy paper emphasizes navigation architecture rather than a new hardware design (Wang et al., 22 Aug 2025). Across both formulations, the central technical issue is that extreme steering freedom and redundant actuation enlarge the maneuver set but complicate modeling, motion-mode definition, actuator coordination, and physically feasible control.

1. Definition, chassis architecture, and actuation redundancy

The defining property of an AWOISV is independent steering at all wheel corners over $(-\pi/2,\pi/2)$, typically implemented with E-corner modules in which steering and driving are integrated at each wheel corner (Yang et al., 19 Aug 2025). This steering freedom distinguishes AWOISV from more conventional 4WS or AWISV configurations that assume front/rear steering with small steering angles and Ackermann-like geometry. The resulting maneuver envelope includes zero-radius turning, diagonal motion, lateral motion, yaw-in-place-like motion, and arbitrary combinations of translation and rotation rather than only conventional normal-steering behavior (Yang et al., 19 Aug 2025).

The associated 4WISD AMR formulation makes the same structural point in wheel-level variables. In the local frame $xoy$, the geometric parameters are the wheelbase $L$, steering track $W$, wheel steering angles

$$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$

and wheel velocities

$v_{fl}, v_{fr}, v_{rl}, v_{rr}.$

Because the platform has more actuation variables than task-space degrees of freedom, the inverse mapping from desired motion to wheel commands is underdetermined (Wang et al., 22 Aug 2025).

A common simplification is to treat such vehicles as if they were fully holonomic in an unrestricted actuator-level sense. The available evidence does not support that simplification. The 4WISD navigation framework explicitly supports “omnidirectional motion and zero-radius turns,” but does not claim a fully holonomic ideal model in the strict robotics sense; instead, it enforces quasi-omnidirectional and mode-constrained motion to preserve physical feasibility (Wang et al., 22 Aug 2025). This suggests that AWOISV capability is best understood as a combination of unusually broad motion authority and nontrivial geometric, tire, and actuator constraints.

2. Motion geometry, instantaneous center of rotation, and mode structure

A major contribution of the AWOISV modeling literature is the replacement of direct instantaneous-center-of-rotation (ICR) coordinates $(B_0,C_0)$ with the theoretical steering radius angle and sideslip angle $(\theta_R,\beta_R)$ representation (Yang et al., 19 Aug 2025). For the $\pm 90^\circ$0-th axle wheel,

$\pm 90^\circ$1

with special handling when $\pm 90^\circ$2. The mapping to bounded angular variables is

$\pm 90^\circ$3

with $\pm 90^\circ$4 and $\pm 90^\circ$5 (Yang et al., 19 Aug 2025). The technical significance is that ICR location, previously represented in an unbounded Cartesian form, becomes a pair of bounded control-like variables suitable for motion-mode analysis and MPC design.

Within this representation, the paper defines six motion modes: longitudinal straight driving mode (LoSDM), diagonal straight driving mode (DSDM), lateral straight driving mode (LaSDM), longitudinal steering mode (LoSM), lateral steering mode (LaSM), and pivot steering mode (PSM). The switching logic depends on the relative position of the ICR to wheel-track and axle geometry. In particular, if $\pm 90^\circ$6, the vehicle is in straight or diagonal motion; if $\pm 90^\circ$7, it is in straight-driving-type modes; if $\pm 90^\circ$8, motion belongs to longitudinal steering; and if $\pm 90^\circ$9, it belongs to lateral steering or pivot steering depending on $(-\pi/2,\pi/2)$0 relative to the axle envelope (Yang et al., 19 Aug 2025).

The 4WISD autonomy framework uses a different but related mode set at the low level: Steering Mode (SM), Oblique Mode (OM), Lateral Mode (LM), and Rotation Mode (RM). These modes are not a rival kinematic theory; rather, they are a feasibility-enforcement mechanism that resolves redundant actuation by selecting structured wheel configurations (Wang et al., 22 Aug 2025).

Framework	Modes	Functional role
$(-\pi/2,\pi/2)$1-$(-\pi/2,\pi/2)$2 representation	LoSDM, DSDM, LaSDM, LoSM, LaSM, PSM	Defines motion modes and switching criteria
Fuzzy low-level 4WISD control	SM, OM, LM, RM	Selects kinematically valid wheel configurations

The two taxonomies are consistent in one important respect: they reject the idea that “omnidirectional” motion eliminates the need for mode reasoning. The AWOISV model shows that some mode families transition continuously whereas others generally require the vehicle to stop. Specifically, LoSDM, DSDM, and LoSM can transition continuously; LaSDM, LaSM, and PSM can transition continuously; transitions between the two groups generally require the vehicle to stop (Yang et al., 19 Aug 2025). The 4WISD controller reaches a similar practical conclusion by forcing commands into one of four feasible modes instead of allowing arbitrary actuator-level combinations (Wang et al., 22 Aug 2025).

3. Kinematics, dynamics, and tire-force modeling

The 4WISD navigation paper separates macro kinematics for navigation from micro kinematics for wheel-level actuation. At the macro level, the robot pose in the global frame is $(-\pi/2,\pi/2)$3, and the local-frame velocity command $(-\pi/2,\pi/2)$4 is mapped by

$(-\pi/2,\pi/2)$5

At the micro level,

$(-\pi/2,\pi/2)$6

where $(-\pi/2,\pi/2)$7 depends on $(-\pi/2,\pi/2)$8, $(-\pi/2,\pi/2)$9, and the steering angles. Because four wheel speeds are used to realize three task-space variables, the mapping is underdetermined and cannot be directly inverted without additional structure (Wang et al., 22 Aug 2025).

The AWOISV modeling paper addresses a related problem from vehicle dynamics rather than mobile-robot navigation. It critiques the traditional single-track $xoy$0-$xoy$1-$xoy$2 model because it assumes small wheel angles and small sideslip angle, using the approximation

$xoy$3

which fails in special modes with large steering angles and large lateral velocity (Yang et al., 19 Aug 2025). The paper instead derives full-body dynamics

$xoy$4

$xoy$5

$xoy$6

together with wheel rotational dynamics

$xoy$7

For lateral tire force, the same work uses an improved Fiala brush model with saturation,

$xoy$8

where

$xoy$9

It then transforms from $L$0 to the generalized $L$1 state using

$L$2

so that $L$3 is forward speed magnitude, $L$4 is vehicle sideslip angle, and $L$5 is yaw rate (Yang et al., 19 Aug 2025).

Taken together, these formulations show why AWOISV control is structurally different from ordinary all-wheel steering. The core issue is not merely four steerable wheels; it is the interaction among wheel orientations, lateral tire forces, underdetermined kinematics, and mode-dependent feasibility. A plausible implication is that control designs that ignore these couplings will be reliable only in restricted operating regions.

4. Hierarchical autonomous navigation in AWOISV-like 4WISD systems

The hierarchical navigation framework proposed for 4WISD systems divides decision-making into a high-level Soft Actor-Critic (SAC) policy and a low-level fuzzy inference system (Wang et al., 22 Aug 2025). The SAC policy generates desired motion commands in task space,

$L$6

rather than directly outputting wheel angles and wheel velocities. This separation is central to the framework: DRL is used for global decision-making and adaptation, while the fuzzy layer guarantees mode consistency and kinematic validity.

The state is given as

$L$7

with sensor state defined by a dual-LiDAR scan with 36 normalized samples,

$L$8

goal state

$L$9

and inner state from the IMU,

$W$0

The reward is

$W$1

with progress reward proportional to $W$2, a safety penalty activated when $W$3, and a stability penalty

$W$4

This reward explicitly discourages unstable or jerky motion, which the paper identifies as especially harmful for omnidirectional wheel platforms (Wang et al., 22 Aug 2025).

The SAC implementation uses an actor with two hidden layers of 512 ReLU neurons, producing mean and standard deviation for a stochastic action distribution; the critic has a similar fully connected architecture and outputs a scalar $W$5-value. Entropy regularization is included in the policy objective, two critics are used to reduce overestimation bias, and target networks are soft-updated (Wang et al., 22 Aug 2025). The paper reports that the constrained hierarchy improves exploration quality and stabilizes training compared with pure DRL.

The low-level fuzzy controller converts $W$6 into a motion mode and corresponding wheel configuration. Its three linguistic inputs are turning radius,

$W$7

velocity offset angle,

$W$8

and energy ratio,

$W$9

with

$$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$0

The output is the motion mode $$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$1. Membership functions are triangular and trapezoidal, the rule base is two-level, Mamdani inference is used, and defuzzification uses Mean of Maximum (MOM) (Wang et al., 22 Aug 2025).

The practical role of the fuzzy layer is to prevent arbitrary wheel motions that would induce wheel dragging, slip, excessive friction, mechanical strain, and unsafe steering-angle combinations. In this framework, “omnidirectional” behavior is realized as a high-level task-space intent that is subsequently projected onto a feasible motion mode rather than as unconstrained actuator-level freedom (Wang et al., 22 Aug 2025).

5. FT-LTVMPC and simultaneous tracking of lateral position and arbitrary heading

For path tracking, the AWOISV control paper proposes a filtered tube-based linear time-varying MPC (FT-LTVMPC) strategy built around the $$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$2-$$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$3 representation (Yang et al., 19 Aug 2025). The control architecture comprises four modules: preprocessing and reference path generation, solving for $$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$4, converting $$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$5 to wheel angles, and forward velocity control. The path-tracking problem is concentrated in the $$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$6-$$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$7 solving module.

The prediction model is written in Frenet coordinates with

$$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$8

After linearization around the current operating point,

$$\delta_{fl}, \delta_{fr}, \delta_{rl}, \delta_{rr},$$9

and discretization gives

$v_{fl}, v_{fr}, v_{rl}, v_{rr}.$0

where $v_{fl}, v_{fr}, v_{rl}, v_{rr}.$1 represents bounded disturbances or model uncertainty (Yang et al., 19 Aug 2025). The MPC cost penalizes lateral deviation, heading deviation, control effort, and terminal-state deviation: $v_{fl}, v_{fr}, v_{rl}, v_{rr}.$2

The robustness mechanism combines tube MPC with filtering. A nominal model is optimized,

$v_{fl}, v_{fr}, v_{rl}, v_{rr}.$3

the actual state error is

$v_{fl}, v_{fr}, v_{rl}, v_{rr}.$4

and the applied control is

$v_{fl}, v_{fr}, v_{rl}, v_{rr}.$5

The error dynamics are

$v_{fl}, v_{fr}, v_{rl}, v_{rr}.$6

A Kalman filter estimates $v_{fl}, v_{fr}, v_{rl}, v_{rr}.$7, and dynamic hysteresis suppresses excessive steering updates due to tiny error fluctuations. The robust positively invariant set $v_{fl}, v_{fr}, v_{rl}, v_{rr}.$8 is then used to tighten constraints through the Pontryagin difference (Yang et al., 19 Aug 2025).

The controller enforces bounds on $v_{fl}, v_{fr}, v_{rl}, v_{rr}.$9, $(B_0,C_0)$0, $(B_0,C_0)$1, and $(B_0,C_0)$2, a terminal set constraint, ICR-position constraints, and wheel steering-rate constraints

$(B_0,C_0)$3

This formulation is designed to achieve simultaneous tracking of lateral deviation $(B_0,C_0)$4 and heading offset $(B_0,C_0)$5, with $(B_0,C_0)$6 not restricted to the path tangent (Yang et al., 19 Aug 2025). That point is fundamental: the controller exploits the extra degrees of freedom of AWOISV so that position tracking and heading tracking are not collapsed into a single conventional path-following objective.

6. Experimental validation, industrial deployment, and open technical issues

The 4WISD hierarchical framework is evaluated in simulation against DWA and TEB, both combined with the same fuzzy low-level controller for fairness. The reported metrics are pose precision (PP), average speed (AS), path efficiency (PE), and success rate (SR). The proposed method achieves the highest average speed in all scenarios, the highest success rate including 100% in S2, lower path efficiency than DWA and TEB, and pose precision that is competitive and always within the training tolerance (Wang et al., 22 Aug 2025). The qualitative interpretation given by the results is that the system prioritizes safe, robust, and physically valid execution over shortest-path optimality.

The ablation study compares the hierarchy with a pure DRL method in which SAC directly outputs wheel velocities and steering angles. Pure DRL shows Q-value overestimation early in training and exhibits erratic wheel motions, sharp turns, dragging, and excess friction, whereas the hierarchical method converges faster and is more stable (Wang et al., 22 Aug 2025). This addresses a recurrent misconception that more expressive end-to-end learning is necessarily better suited to redundant omnidirectional chassis. In the reported experiments, constraining the action realization improves both learning behavior and mechanical plausibility.

Real-world validation of the same framework is performed on a custom 4WISD AMR in an active factory with moving machinery, human workers, and predefined pathways. No additional fine-tuning is needed; the model size is 3.5 MB; inference time is $(B_0,C_0)$7 ms/step; fuzzy controller time is $(B_0,C_0)$8 ms/step; and the total system meets the 10 Hz control requirement. For long-distance navigation, $(B_0,C_0)$9 provides global waypoints, while the learned policy handles local motion to each waypoint (Wang et al., 22 Aug 2025). This establishes a concrete industrial deployment pattern: global planning, learned local navigation, and low-level feasibility enforcement.

The FT-LTVMPC framework is validated by TruckSim-MATLAB/Simulink co-simulation and hardware-in-the-loop testing on a 4-axle AWOISV with a control period of 0.02 s, MPC sample time of 0.25 s, and prediction horizon of 20 (Yang et al., 19 Aug 2025). On a sinusoidal path with curvature radius varying from $(\theta_R,\beta_R)$0 m, heading target varying from $(\theta_R,\beta_R)$1, and speeds of 2, 4, 6, and 8 m/s, heading profile tracking remains accurate over all speeds; at 8 m/s, lateral error remains within about 0.15 m and heading error within about $(\theta_R,\beta_R)$2; sideslip angle stays near the theoretical $(\theta_R,\beta_R)$3; and yaw rate remains within constraints (Yang et al., 19 Aug 2025).

Comparative evaluation against C-MPC, NMPC, LTVMPC, T-LTVMPC, and FT-LTVMPC shows that C-MPC performs worst because the single-track normal-steering model cannot represent AWOISV special motion modes; NMPC and LTVMPC improve performance; and T-LTVMPC and FT-LTVMPC are best among the tested methods, with tube-based methods providing clear robustness benefits. Solving time for all but NMPC stays under 20 ms, satisfying real-time needs (Yang et al., 19 Aug 2025). In HIL tests on a composite path combining LoSM, LoSDM, and DSDM with speed varying between 0 and 8 m/s, all three tested methods are real-time feasible; T-LTVMPC reduces median lateral error by 73.9% and heading error by 34.8% relative to LTVMPC; FT-LTVMPC reduces median lateral error by 61.3% and heading error by 37.9%; and the average sliding standard deviation of wheel angles is 0.0407 for LTVMPC, 0.0782 for T-LTVMPC, and 0.0341 for FT-LTVMPC, implying a 56.4% smoothness improvement over T-LTVMPC (Yang et al., 19 Aug 2025).

The main open issues identified in the navigation work are improve reward shaping, add more sensors, address transient misalignment and inter-wheel dragging during mode transitions, and extend to multi-robot coordination for collaborative industrial tasks (Wang et al., 22 Aug 2025). From the AWOISV modeling side, the explicit statement that transitions between longitudinal and lateral mode families generally require the vehicle to stop underscores a separate limitation: broad maneuverability does not eliminate geometric discontinuities in mode switching (Yang et al., 19 Aug 2025). Together, these results suggest that the principal research challenge is no longer whether AWOISV can generate exotic motion primitives, but how to realize those primitives with mode-aware, uncertainty-robust, real-time control that remains mechanically safe in complex operating environments.