Unleashing the Agility of Wheeled-Legged Robots for High-Dynamic Reflexive Obstacle Evasion

Abstract: Wheeled-legged robots combine the energy efficiency of wheeled locomotion with the terrain adaptability of legged systems, making them promising platforms for agile mobility in complex and dynamic environments. However, enabling high-dynamic reflexive evasion against fast-moving obstacles remains challenging due to the hybrid morphology, mode coupling, and non-holonomic constraints of such platforms. In this work, we propose AWARE, Adaptive Wheeled-Legged Avoidance and Reflexive Evasion, a hierarchical reinforcement learning framework for high-dynamic obstacle avoidance in wheeled-legged robots. The proposed system naturally exhibits diverse emergent gaits and evasive behaviors, including forward lunge and lateral dodge, thereby leveraging the robot's hybrid morphology to enhance agility under highly dynamic threats. Extensive experiments in Isaac Lab simulation and real-world deployment on the M20 platform across diverse dynamic scenarios demonstrate that AWARE achieves robust and agile obstacle avoidance while revealing behaviorally distinct evasive strategies. These results highlight both the practical effectiveness of AWARE and the intrinsic reflexive agility of wheeled-legged robots.

• The paper introduces AWARE, a hierarchical reinforcement learning framework that separates high-level decision-making from low-level locomotion to achieve rapid reflexive obstacle evasion.
• It demonstrates robust performance in both simulation and real-world tests, outperforming baseline controllers with nearly 30% success under sub-second reaction times.
• Ablation studies confirm that dual-mode control and curriculum training are critical for managing hybrid constraints and ensuring reliable safety and maneuver efficiency.

High-Dynamic Reflexive Evasion in Wheeled-Legged Robots via Hierarchical Reinforcement Learning

Introduction and Motivation

Wheeled-legged robots combine rolling efficiency with the terrain adaptability of legged systems, providing a unique platform for agile mobility in both structured and unstructured environments. However, the challenge of high-dynamic reflexive obstacle evasion remains unresolved for these robots, primarily due to their hybrid morphology, mode coupling, and non-holonomic constraints. The paper introduces AWARE (Adaptive Wheeled-Legged Avoidance and Reflexive Evasion), a hierarchical reinforcement learning framework specifically designed to address the demands of robust, real-time high-dynamic obstacle avoidance and reflexive evasion for wheeled-legged systems (2604.23761).

Figure 1: AWARE enables agile and robust reflexive evasion and navigation avoidance in wheeled-legged robots experiencing diverse threat scenarios with seamless inter-modal transitions.

Technical Approach

Hierarchical Policy Architecture

AWARE implements a hierarchical structure that distinctly separates high-level decision making from low-level locomotion control. The high-level evasive policy receives proprioceptive and exteroceptive information, most notably the Relative Approaching Rate (RAR), and selects from two pre-trained expert policies:

• Low-Speed Omnidirectional Network: Handles smooth, energy-efficient navigation avoidance for low-threat/distant obstacles.
• High-Dynamic Agile Network: Executes aggressive, reflexive evasion such as forward lunges and lateral dodges under imminent hazard, optimizing for actuation limits and minimal reaction time.

The decision process leverages a Gumbel-Softmax-based mode selector to discretely route commands to the optimal low-level network. This dual-mode architecture explicitly mitigates mode mixing and enables clean, reliable transitions between modalities.

Training Paradigm

AWARE is trained using a decoupled two-stage curriculum:

• Stage 1: Each low-level expert is trained independently to specialize in either high-fluency velocity tracking or agile start-up maneuvers. Training data distributions for the two experts are non-overlapping in acceleration space to encourage specialization.
• Stage 2: The high-level policy is trained atop the frozen experts through curriculum learning, with obstacle speed and threat regularization to suppress over-evasion and false positives.

Domain randomization is extensively used to ensure hardware transferability, randomizing dynamic parameters (actuation, base pose, obstacle approach) for robust sim-to-real performance.

Experimental Results and Empirical Analysis

Obstacle Evasion Strategies and Behavior

AWARE autonomously discovers a spectrum of evasion maneuvers exploiting rolling, stepping, and hybrid gaits as dictated by obstacle speed, approach vector, and available reaction time.

Figure 3: Visualization of dual-mode system; high-speed threats elicit extreme lateral and forward evasion, while low-speed threats are handled by smooth gait transitions.

The system leverages its morpho-dynamic repertoire, locking wheels instantaneously for high-friction deceleration in rolling mode and executing explosive stepping for lateral threats. The policy demonstrates effective threat calibration, only invoking high-dynamic maneuvers when strictly necessary.

Numerical Performance

Empirical evaluation in simulation shows a marked improvement over baseline approaches (REBot and Relaxed-Barrier-MPC), particularly at short reaction times (≤1 s):

Figure 5: AWARE consistently outperforms REBot and RB-MPC in both success rate and spatial efficiency, especially under challenging, high-speed obstacle scenarios.

Reaction Time (s)	ASR (Success Rate)	AMD (Maneuver Distance, m)
0–1	0.294 ± 0.315	0.751 ± 0.246
1–2	0.931 ± 0.058	0.626 ± 0.129
2–3	0.973 ± 0.027	0.580 ± 0.041

For short reaction time intervals (0–1 s), AWARE still retains nearly 30% success—a domain where classical MPC and non-hybrid DRL models fail entirely. This evidences AWARE's capacity for fast, torque-intensive recovery maneuvers.

Behavioral Mode Analysis

t-SNE analysis of joint-space and kinematic features confirms that AWARE yields behaviorally separable motion regimes, cleanly segmenting navigation, hybrid, and reflexive evasion gaits.

Figure 2: t-SNE embedding shows mode clustering; high-speed evasive responses distinct from low-speed navigation.

Force and power profiling indicate that rolling-based start-up strategies achieve higher Transient Escape Acceleration (TEA) and Torque-Normalized Acceleration (TNA) compared to stepping, highlighting the energetic benefit of exploiting the wheeled substructure for on-the-fly lateral or forward movement.

Ablation and Ablative Importance

Ablation studies establish that:

• Removal of dual-mode control dramatically increases mode confusion and decreases safety (catastrophic failure under high threat).
• Omission of curriculum or explicit safety rewards undermines task focus and performance.
• Feasibility regularization is essential for physical realizability of high-level commands.

  Figure 4: Ablative results show all architectural components contribute critically to success and efficiency.

Real-World Deployment and Mixed-Mode Evasion

The framework translates effectively to hardware, with real-robot results demonstrating robust evasion under box throws, kicks, and stick attacks from diverse angles. Importantly, the system autonomously blends navigation avoidance and reflexive evasion, even under abrupt interventions.

Figure 6: Real-robot tests show robust, scenario-adaptive evasive and recovery behaviors with accurate transition between operational modes.

Real-world average avoidance success rate (ASR) is 59% (AMD 1.1 m), with noted degradation due to platform actuation and sim-to-real gap. Nevertheless, the hardware trials validate the policy's agility and transition capacity.

Implications and Future Outlook

AWARE pushes the state of the art by providing a versatile reflexive evasion architecture uniquely suited to the hybrid dynamic constraints of wheeled-legged robots. The findings have several concrete implications:

• Design of Hybrid Mobility Systems: Explicit mode decoupling is necessary to unlock the full reactive agility potential present in hybrid platforms; pure DRL or single-mode MPC frameworks are insufficient.
• Policy Specialization and Routing: Hierarchical DRL with expert routing notably enhances physical feasibility, interpretability, and real-world transfer.
• Sim-to-Real Robotics: Successful domain-randomized deployment paves the way for closed-loop DRL in highly dynamic, human-centric environments beyond simulation.

Potential future directions include scaling the approach to multi-agent and cooperative evasion, online adaptation to unforeseen platform failures, and integration with real-time perception for high-speed outdoor navigation.

Conclusion

This work sets a new reference for high-dynamic obstacle avoidance in wheeled-legged robotics, combining hierarchical reinforcement learning, mode specialization, and curriculum-driven training to realize practical reflexive evasion. Quantitative and qualitative benchmarks on both simulated and hardware platforms confirm the viability and advantages of mode-decoupled architectures in challenging dynamic scenarios.