PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Air-Ground Motion Control with Temporal Mode Switching and Cross-Terrain Tracking

Ruitian Pang, Mingrui Li, Xuanting Liu, Tiancheng Lai, Juncheng Chen, Xiangyu Li, Ruibin Zhang, Qishao Wang, Jin Yu, Haiyin Piao, Fei Gao, Chao Xu, Yanjun Cao

arXiv:2609.26564v1cs.RO

TL;DR

Limited terrain perception makes reliable air-ground switching difficult, and terrain-varying wheel interactions challenge consistent tracking. The paper combines a temporal mode selector with reinforcement-learning trajectory tracking. The integrated system completes a 101 m air-ground trajectory with a position RMSE of 0.08 m.

  • Problem

    Limited onboard perception complicates reliable air-ground switching, while variations in wheel-ground interaction across terrains challenge consistent ground trajectory tracking.

    Source: §I · p. 1

  • Method

    A learned selector uses historical ToF and robot-state information with future reference information, while RL policies use robot observations and reference previews for tracking.

    Source: §I · p. 2

  • Results

    The complete system achieves a position RMSE of 0.08 m over a 101 m real-world air-ground trajectory, while ground tracking reduces Ep relative to PID by 92.2%, 77.9%, and 75.8% on deformable soft, uneven grass, and rough block terrains, respectively.

    Source: 2) Real-World Locomotion Evaluation · p. 6, §V · p. 8

  • Takeaways & Limitations

    Experiments demonstrate robust ground tracking across diverse terrains and reliable air-ground switching with limited onboard perception.

    Source: §V · p. 8

Abstract

from arXiv · show

Passive-wheeled terrestrial-aerial bimodal vehicles (TABVs) combine aerial mobility with energy-efficient ground locomotion. However, reliable air-ground mode switching under limited onboard perception and robust ground trajectory tracking across diverse terrains remain challenging when targeting real-world applications. In this work, we propose a learning-based air-ground motion control framework for passive-wheeled TABVs: 1) a learned mode selector for autonomous air-ground motion mode switching. The selector uses historical single-point time-of-flight (ToF) measurements and robot states together with future reference information to determine the active locomotion mode. 2) a reinforcement learning control policy for trajectory tracking. The policy combines proprioceptive observations with future reference information to anticipate trajectory changes. For ground locomotion, multi-terrain training and dynamics randomization enable robust tracking across different terrains. Simulation and real-world experiments demonstrate reliable air-ground switching under limited perception and accurate ground tracking across diverse terrain conditions. The learned selector outperforms a rule-based mode selector in challenging transitions, while the ground controller achieves lower position RMSE than PID across all tested conditions and maintains decent tracking where NMPC fails. With these capabilities integrated, the system tracks a 101m air-ground trajectory through multiple autonomous mode transitions with a position RMSE of 0.08m.

I. INTRODUCTION

Passive-wheeled TABVs offer energy-efficient ground travel alongside aerial mobility, but limited terrain perception complicates mode switching and terrain-dependent wheel interactions undermine consistent tracking. The proposed framework combines learned switching with trajectory-tracking policies, avoiding dense terrain representations and handcrafted terrain-specific switching rules.

I. INTRODUCTION

  • TABVs combine aerial mobility with energy-efficient ground locomotion, but limited onboard perception complicates reliable switching and terrain-varying wheel interactions challenge consistent ground tracking.Ground locomotion generally uses less energy and produces less acoustic noise than aerial locomotion.Source: §I · p. 1

A. Mode Switching in Bimodal Robots

I. INTRODUCTION

  • The paper proposes an integrated framework for autonomous switching and aerial and ground trajectory tracking in passive-wheeled TABVs.Source: §I · pp. 1–2
  • A dual-branch selector uses historical single-point ToF measurements, robot states, and future reference information for online air-ground switching without terrain-specific rules or dense terrain representations.Ground tracking uses reinforcement learning, local reference preview, multi-terrain training, and dynamics randomization, with zero-shot sim-to-real transfer.Source: §I · p. 2

B. Ground Control of Passive-Wheeled TABVs

Prior ground controllers rely on simplified wheel-ground models or terrain-specific tuning, while cross-terrain learning for passive-wheeled TABVs remains less explored. The framework tracks a time-parameterized reference with separate aerial and ground policies and mode selection.

B. Ground Control of Passive-Wheeled TABVs

III. METHODOLOGY

  • The framework uses separate ground and aerial RL policies for tracking and a dual-branch temporal selector driven by ToF and robot-state histories plus current reference intent.Source: §III · p. 3

A. Problem Formulation and Reference Trajectory

B. Reinforcement Learning Control Policies

Ground and aerial control are modeled as separate reinforcement-learning problems. Their policy observations combine robot-state information and history with future trajectory samples to capture current tracking deviation and upcoming motion.

B. Reinforcement Learning Control Policies

  • Ground and aerial control are modeled as separate MDPs, each with its own state, action, dynamics, reward, and policy.Each policy is optimized to maximize expected discounted return.Source: B. Reinforcement Learning Control… · p. 3

1) Observation Space:

  • Policy observations include robot velocities and orientation, the previous normalized action, and a history of angular velocity, linear velocity, and actions.Source: 1) Observation Space · p. 3
  • The tracking preview samples future reference positions and velocities, providing both current tracking deviation and the reference trajectory’s future motion trend.Source: 1) Observation Space · p. 3

2) Action Space:

Both policies output normalized actions that are converted through a shared CTBR interface. Their reward formulations include tracking objectives and action or motion penalties.

2) Action Space:

  • The normalized four-dimensional action is mapped to CTBR commands by scaling thrust to [Tmin, Tmax] and angular velocities by ωmax.Both policies use this shared interface for mode switching.Source: 2) Action Space · p. 3

3) Reward Function:

  • Ground reward terms address horizontal position and motion-direction speed tracking, with penalties for heading error, excessive pitch rates, large roll angles, and action changes.The cited reward fragments identify the motion-direction speed terms and define heading error, pitch rate, roll angle, and normalized action.Source: 3) Reward Function · pp. 3–4
  • Aerial reward terms address three-dimensional position tracking, heading alignment, and velocity-direction alignment, with an action-smoothness penalty.Source: 3) Reward Function · p. 4

4) Network Architecture:

The aerial and ground policies share an asymmetric actor-critic architecture, with three-layer MLPs using [512, 256, 128] hidden dimensions and ELU activations. During training, actors receive noisy observations while critics use privileged noise-free state information.

4) Network Architecture:

  • Both aerial and ground policies use three-layer actor and critic MLPs with hidden dimensions [512, 256, 128] and ELU activations.They share the same asymmetric actor-critic architecture.Source: 4) Network Architecture · p. 4
  • Actors train on noisy observations, while critics receive privileged noise-free state information for value estimation.Noisy actor inputs are used to improve robustness to sensing uncertainty.Source: 4) Network Architecture · p. 4

C. Mode Switching via a Dual-Branch Temporal Network

The selector classifies the active air-ground mode using robot-state history and current reference intent. Its dual-branch network combines temporal ToF and state features with a reference preview, then applies confirmed, asymmetric switching thresholds.

C. Mode Switching via a Dual-Branch Temporal Network

  • Air-ground mode selection is formulated as binary classification from a state-history branch and a reference-intent branch.The selector uses historical single-point ToF measurements and robot states alongside future reference information.Source: C. Mode Switching via a Dual-Branch… · p. 4
  • The state input includes normalized single-point ToF and robot-state features, with the state branch processing a sequence of past inputs.Valid ToF returns from 0.03–20 m are mapped to [0, 1], and invalid measurements are set to zero.Source: C. Mode Switching via a Dual-Branch… · p. 4
  • Reference intent encodes ten future relative reference-height changes and corresponding vertical velocities to characterize the trajectory’s future vertical evolution.The relative height changes compare future reference heights with the current reference height.Source: C. Mode Switching via a Dual-Branch… · p. 4
  • A two-layer GRU processes state history while a two-layer MLP encodes the current reference preview; concatenated features yield Ground and Air logits and mode probabilities.Only the state branch uses temporal history, keeping the reference-intent branch current and responsive to the latest motion command.Source: C. Mode Switching via a Dual-Branch… · p. 4
  • Switching requires consecutive-frame confirmation: Air probability ≥0.50 for two frames triggers Ground-to-Air, while ≤0.20 for eight frames triggers Air-to-Ground.These asymmetric thresholds are used in simulation and real-world experiments to suppress mode chattering while favoring faster takeoff transitions and conservative returns to Ground.Source: C. Mode Switching via a Dual-Branch… · p. 4
  • The selector is supervised from Isaac Lab rollouts labeled by scenario-specific rules for Takeoff, Ground Support Loss, Landing, and Ground Recovery.Frozen aerial or ground experts track the reference according to each target mode; ToF inputs are perturbed during data collection, and the selector is trained with cross-entropy loss.Source: C. Mode Switching via a Dual-Branch… · pp. 4–5

A. Experimental Setup

Real-world tests compare ground tracking across four terrains and evaluate hybrid air-ground tracking. The trained policies track across unseen and differing terrains, while the mode selector combines independently trained aerial and ground policies through a shared interface.

A. Experimental Setup

  • Real-world experiments use onboard ToF with odometry and IMU state estimates; policies run at 200 Hz and the mode selector at 50 Hz.The physical platform runs online inference and control on an NVIDIA Jetson NX.Source: A. Experimental Setup · p. 5

1) Training Configuration:

  • Ground-policy training uses flat, continuous rough, discrete rough, and terrain-transition regions, with proportions of 50%, 25%, 20%, and 5%, respectively, and randomizes dynamics to improve sim-to-real transfer.Randomized parameters include mass, wheel-ground friction, actuation delays, motor response, observation noise, and external disturbances.Source: 1) Training Configuration · p. 5
  • Training trajectories include ground reference speeds up to 1.9 m/s on flat and 1.0 m/s on rough terrain, while aerial speeds range from 0.8–2.0 m/s at heights of 1.0–3.0 m.The ground and aerial policies are trained using PPO with 4096 parallel environments and require approximately 40 minutes each.Source: 1) Training Configuration · p. 5

2) Real-World Locomotion Evaluation:

  • Ground controllers are compared on flat rigid, deformable soft, uneven grass, and rough-block terrains using the same figure-eight reference, fixed parameters across terrains, and no explicit terrain identification for the proposed method.The proposed method is evaluated in real-world multi-terrain ground tracking and air-ground hybrid trajectory tracking.Source: 2) Real-World Locomotion Evaluation · pp. 5–6
  • Compared with NMPC, the proposed method reduces Ep by 33.9% on flat rigid terrain and 17.1% on deformable soft terrain; compared with PID, it reduces Ep by 92.2%, 77.9%, and 75.8% on soft, grass, and rough-block terrain, respectively.The deformable soft and uneven grass terrains are unseen during training, and reference speeds on all three non-flat terrains exceed the 1.0 m/s maximum used for rough-terrain training.Source: 2) Real-World Locomotion Evaluation · p. 6
  • Over a hybrid trajectory spanning ground locomotion, takeoff, aerial flight, and landing, the policies achieve position RMSE Ep = 0.097 m and velocity RMSE Ev = 0.252 m/s.The reported errors cover the complete trajectory.Source: 2) Real-World Locomotion Evaluation · p. 6, Figure 6 · p. 6
  • Across all eight tested conditions, the proposed method achieves lower position RMSE (Ep) than PID, while NMPC fails to complete trajectories on uneven grass and rough block at both tested speeds.Table I summarizes four conditions shown in Fig. 5 and four additional lower-speed conditions.Source: Table I · p. 6, 2) Real-World Locomotion Evaluation · p. 6
  • Independently trained ground and aerial policies combine through the shared CTBR interface without additional transition-specific control design.Source: 2) Real-World Locomotion Evaluation · p. 6

C. Temporal Air-Ground Mode Selector Evaluation

The selector is evaluated in simulation across reference-driven and support-change transitions, with training designed to account for sensing uncertainty. It achieves high switching success in the reported scenarios, including rapid response to support loss.

1) Training Configuration:

  • Training used 4096 parallel Isaac Lab environments and ToF perturbations including dropout, noise, sequence-wise bias, and 0–1-frame latency.The dataset was split by environment 8:1:1; motion-state dropout probability was 0.10.Source: 1) Training Configuration · p. 6

2) Simulation Switching Performance:

  • The simulation evaluation covered four scenarios, including reference-driven Takeoff and Landing and support-triggered Ground Support Loss and Ground Recovery transitions.Switching delay is measured as actual policy activation time minus the scenario-specific label transition time; negative values indicate early switching. A successful trial must complete the transition within the evaluation horizon without an incorrect transition, with an additional height constraint for Ground Support Loss.Source: 2) Simulation Switching Performance · p. 6
  • 100% success was reported in Takeoff, Landing, and Ground Recovery, with mean switching delays of 126 ms, 125 ms, and 262 ms, respectively.Table II reports switching delays as mean ± standard deviation.Source: 2) Simulation Switching Performance · pp. 6–7, Table II · p. 7
  • Across 250 Ground Support Loss trials, the selector achieved 100% success with a mean switching delay of 28 ± 12 ms.The test covered five terrain-drop heights from 0.075 to 0.275 m, with 50 trials at each height.Source: 2) Simulation Switching Performance · p. 7

3) Mode Selector Ablation Study:

The ablation study tests the contributions of ToF history and future reference intent to mode selection. Removing ToF sharply reduces success in support-loss trials, while removing intent eliminates Takeoff success.

3) Mode Selector Ablation Study:

  • The Full selector uses 20-frame ToF and motion-state histories with future reference intent; ablations remove ToF, intent, or historical ToF while retaining other inputs as specified.All variants share the same architecture, data split, and training configuration and are trained from scratch.Source: 3) Mode Selector Ablation Study · p. 7
  • Removing ToF reduced Ground Support Loss success from 100% to 3%, while removing future intent reduced Takeoff success from 100% to 0%.The ablations were evaluated over 100 trials per model per scenario.Source: Table IV · p. 7, 3) Mode Selector Ablation Study · p. 7
  • Using only current ToF also lowered Ground Support Loss and Ground Recovery success relative to Full, indicating a benefit from temporal ToF information.Switching delays are averaged only over successful trials and should be read alongside success rates.Source: 3) Mode Selector Ablation Study · p. 7

4) Comparison with a Rule-based mode selector:

The learned selector outperforms the rule-based selector in two challenging simulated transitions, and the trained selector also switches modes on the physical robot. Integrated real-world testing reports a 101 m air-ground trajectory with 0.08 m position RMSE.

4) Comparison with a Rule-based mode selector:

  • The learned selector achieved 92.0% and 96.0% success in Gradual Support Loss and Fast Narrow Gap, versus 54.0% and 56.0% for the rule-based selector.Both methods used the same ToF stream, 50 Hz decision frequency, two-frame confirmation, simulation initializations, and ToF perturbations.Source: 4) Comparison with a Rule-based mode… · p. 7, Table V · p. 8

5) Real-World Mode Switching Evaluation:

D. Integrated System Evaluation

  • The integrated system completed a 101 m real-world air-ground trajectory with multiple autonomous mode transitions and 0.08 m position RMSE.The conclusion describes robust ground tracking across diverse terrains and reliable mode switching.Source: D. Integrated System Evaluation · p. 8, §V · p. 8

V. CONCLUSION

  • The framework combines independently trained aerial and ground policies with a temporal selector for autonomous switching under limited onboard perception.The conclusion identifies more complex and dynamic environments as future work.Source: §V · p. 8

Related papers

Loading 2609.26564v1…