PaperlayerView exact revision on arXiv

Source-linked AI summary

CeDR-Nav: Cross-Embodiment Mapless Navigation via Offline Demonstration Recasting

Xianghui Wang, Xinming Zhang, Yiwei Qian, Siwei Cheng, Shanze Wang, Dan Zhang, Wei Zhang

arXiv:2610.13007v1cs.RO

TL;DR

Source demonstrations may contain reusable navigation experience, but embodiment mismatch prevents their direct use as valid target supervision. CeDR-Nav recasts those demonstrations and trains a dedicated policy for each target without online RL training; in real-world tests, it achieved a 100% success rate for both goals across four TurtleBot2 configurations.

  • Problem

    Source demonstrations contain potentially reusable navigation experience, but embodiment geometry and velocity limits prevent them from directly serving as valid target supervision.

    Source: §I · p. 1

  • Method

    CeDR-Nav uses geometry-consistent scaling and validity assessment to recast source demonstrations, then trains a target-specific policy using a Gaussian mixture model.

    Source: §I · p. 1

  • Results

    100% success rate for both goals across all four TurtleBot2 configurations in real-world tests.

    Source: C. Real-World Deployment Results · p. 8

  • Takeaways & Limitations

    Target-specific policies can be derived from single-source-embodiment demonstrations without additional demonstration collection or online RL training, with training taking seconds.

    Source: §VII · p. 8

Abstract

from arXiv · show

Learning-based methods have demonstrated strong performance in mapless robot navigation. However, a learned policy remains restricted to the embodiment on which it was trained. Many cross-embodiment navigation methods require time-consuming online training and can still underperform target-specific policies. In fact, demonstrations collected on a source embodiment already contain navigation experience that is potentially reusable for deriving a policy for each target embodiment, but the embodiment mismatch prevents these demonstrations from directly serving as valid supervision for the target. In this article, we present CeDR-Nav, a cross-embodiment demonstration recasting framework that trains a dedicated navigation policy for each target embodiment from existing source demonstrations without online data collection. Geometry-consistent scaling with per-demonstration validity assessment converts the source demonstrations into target training data, and a target-specific Gaussian mixture policy generates feasible avoidance maneuvers for each target embodiment. Extensive quantitative simulations and real-world deployments across multiple robot platforms demonstrate the effectiveness of CeDR-Nav, with an average success-rate gain of 13.8 percentage points over DRL-DCLP in cross-embodiment simulations and target-specific policy training completed within seconds.

I. INTRODUCTION

Learned navigation policies are commonly embodiment-specific, while source demonstrations could be reused but need conversion into valid target supervision. CeDR-Nav learns dedicated target policies offline, addressing limitations in transfer, training coverage, and data preparation.

I. INTRODUCTION

  • Source demonstrations contain potentially reusable navigation experience, but embodiment geometry and velocity limits prevent their direct use as valid target supervision.Source: §I · p. 1
  • CeDR-Nav recasts existing source demonstrations into target-valid data and trains a dedicated policy for each target without online RL training.Source: §I · pp. 1–2
  • Simulations and real-world experiments across multiple platforms and tasks demonstrate effective navigation policies derived from source demonstrations.Source: §I · p. 2

II. RELATED WORK

  • Policy-transfer and universal-policy approaches can be limited by conservative transfer, training demands, or performance outside the training distribution; CE-Nav also requires online refinement for each target.Source: §II · p. 2
  • Data-driven approaches rely on multi-platform datasets or scene-geometry estimation, while CeDR-Nav instead learns target policies from existing demonstrations without additional online interaction.Source: §II · p. 2

III. PRELIMINARIES

The paper formulates mapless navigation as imitation of an expert controller using a dataset of successful episodes. It specifies the robot state, expert action, and embodiment parameters used in that formulation.

III. PRELIMINARIES

  • The policy πθ is trained on expert demonstrations by minimizing a loss between expert actions and policy outputs.Source: §III · p. 2
  • Each demonstration is a successful H-step episode, and the dataset comprises J such episodes.Source: §III · p. 2
  • The state contains laser scans, relative goal position, and current velocities; actions are linear and angular velocity commands, while embodiment parameters specify footprint extents and speed limits.Source: §III · p. 2

IV. METHODOLOGY

CeDR-Nav first recasts source demonstrations into target-valid training data, then trains a target-specific policy. Recasting scales observations and actions while preserving geometric consistency and respecting target velocity limits.

IV. METHODOLOGY

  • The methodology has two phases: recast source demonstrations into valid target data, then train the target-specific policy on those demonstrations.Source: §IV · p. 2

1) Geometry-Consistent Scaling:

  • Recasting scales distance components by λ while leaving angles unchanged, keeping observations and actions geometrically consistent with the source demonstration.Source: 1) Geometry-Consistent Scaling · p. 3
  • When the ideal recast action exceeds target velocity limits, CeDR-Nav scales both velocity components by β to satisfy the limits while preserving turning radius.Source: 1) Geometry-Consistent Scaling · p. 3

2) Validity Assessment:

Validity assessment checks recast demonstrations against the target footprint and selects the smallest passing candidate scale for each demonstration. Demonstrations with no passing scale are excluded.

2) Validity Assessment:

  • Collision assessment flags a demonstration when a recorded recast scan endpoint falls inside the target footprint; otherwise it is collision-free.Source: 2) Validity Assessment · p. 3
  • The assessment checks recorded scan endpoints, not the continuous swept volume between successive states; the reference scale guarantees validity for source-valid demonstrations, while a lower scale bound can exclude demonstrations needing more scaling.Source: 2) Validity Assessment · pp. 3–4
  • For each demonstration, CeDR-Nav selects the smallest candidate scale that passes validity assessment and excludes the demonstration if none passes.Source: 2) Validity Assessment · p. 3

B. Target-Specific Policy Learning

CeDR-Nav learns a dedicated policy for each target embodiment from recast demonstrations, using a Gaussian mixture to model multimodal actions and producing a single velocity command at deployment.

B. Target-Specific Policy Learning

A. Experimental Setup

The experiments use source demonstrations collected in ROS Stage and train separate policies for five target configurations.

  • The source demonstrations were collected in a cluttered 10 × 10 m2 ROS Stage room using MAER-Nav.Source: A. Experimental Setup · p. 4
  • The rectangular differential-drive source platform used velocity limits of (0.5 m/s, π/2 rad/s), and successful-episode filtering yielded 12,047 episodes with approximately 2.1 million state–action pairs.Source: A. Experimental Setup · p. 4
  • Five ROS Stage targets used velocity limits of (2.0 m/s, π/2 rad/s); independent recasting admitted 8,572–8,955 episodes per target, or 71–74%.Source: A. Experimental Setup · p. 4
  • Training used Adam with a 3 × 10−4 learning rate, batch size 512, and 150,000 gradient updates.Source: A. Experimental Setup · p. 5

B. Cross-Embodiment Evaluation

Across five target configurations, CeDR-Nav outperforms DRL-DCLP in success rate in two unseen simulation scenarios; on BARN, it also achieves a higher score with higher success and lower collision rates.

B. Cross-Embodiment Evaluation

  • The simulation evaluation tested five target configurations in unseen cluttered scenarios SEnv1 and SEnv2, using identical maps for both methods under each configuration.Three random-seed policies per target were each evaluated on 100 random tasks per scenario.Source: B. Cross-Embodiment Evaluation · p. 5
  • Success rates rose from 88.6% to 93.0% in SEnv1 and from 73.6% to 96.8% in SEnv2, with CeDR-Nav higher in every configuration.SEnv2 gains were 15–29 percentage points, mainly reflecting fewer timeouts with comparable collision rates.Source: B. Cross-Embodiment Evaluation · p. 5

C. Evaluation on BARN Benchmark

  • The BARN score averages test scores based on collision-free goal completion and actual traversal time relative to optimal traversal time.Source: C. Evaluation on BARN Benchmark · p. 5
  • On BARN, CeDR-Nav scored 0.4313 versus DRL-DCLP's 0.4200, with success rates of 86.33% versus 84.00% and collision rates of 3.00% versus 6.67%.The evaluation used all 300 public BARN maps with shared velocity limits, collision thresholds, and timeout settings.Source: C. Evaluation on BARN Benchmark · p. 5

D. Training Time Analysis

CeDR-Nav trains target policies in seconds, and its recasting and GMM design choices outperform the reported alternatives in SEnv2. Real-world evaluation used three platforms and unseen test scenarios without fine-tuning.

D. Training Time Analysis

  • CeDR-Nav completes 150,000 gradient updates per target in 12.15 s on an RTX 4090, and all five target policies take about 61 s versus 74 h 20 min for one DRL-DCLP policy.It also achieves higher SR than DRL-DCLP on the same configurations; timing excludes JIT compilation and waits for GPU execution to complete.Source: D. Training Time Analysis · p. 5

E. Ablation Studies

  • 96.80% average SR with episode filtering and per-demonstration scale selection compares with 84.20% for fixed λref recasting and 47.87% without recasting.Recasting improves SR across all five configurations; filtering and per-demonstration scale selection together further improve it over conservative full-dataset recasting.Source: E. Ablation Studies · p. 5
  • 96.8% average SR with CeDR-Nav compares with 96.2% when the same admitted episodes are all recast at λM = 1.5.CeDR-Nav selects the smallest candidate scale that passes validity assessment for each admitted demonstration.Source: E. Ablation Studies · p. 5
  • 96.80% average SR with the GMM policy compares with 94.07% for a plain MLP policy without a mixture output.All other settings were kept the same.Source: E. Ablation Studies · p. 5
  • 96.8% average SR with CeDR-Nav compares with 69.07% for Deploy-Scaling, with the largest drops on wider configurations.Deploy-Scaling uses a source policy across all five target configurations without target-specific retraining.Source: E. Ablation Studies · p. 5

VI. REAL-WORLD PERFORMANCE EVALUATION

  • Configuration-specific policies were tested on three platforms in scenarios unseen during training, without fine-tuning or checkpoint selection using real-world data.The supplied passage describes the evaluation setup, not its outcomes.Source: §VI · p. 5

A. Hardware Setup

The experiments used TurtleBot2, Unitree Go2, and Booster K1 configurations with platform-specific sensing and localization. Testing covered matched TurtleBot2 comparisons, Go2 cargo and dynamic scenarios, and a different sensor type on K1.

A. Hardware Setup

  • TurtleBot2 used a Hokuyo 2-D LiDAR and AMCL, Go2 used a Livox 3-D LiDAR converted to a 2-D scan and FAST-LIO2, and K1 used stereodepth-derived 2-D scans and onboard odometry.Relative goals were supplied; maps were not used for motion planning.Source: A. Hardware Setup · p. 6

B. Testing Scenarios and Task Description

  • Figure 6 groups trajectories by platform and method: TurtleBot2 configurations compare CeDR-Nav with DRL-DCLP under the same obstacle layouts, while Go2 and K1 show CeDR-Nav tests.The Go2-C2 trajectory includes the dynamic pedestrian test.Source: Figure 6 · p. 7, B. Testing Scenarios and Task… · p. 7
  • The real-world tests covered seven configurations, including four TurtleBot2 configurations, two Go2 cargo states, and one Booster K1 configuration.The Go2-C2 cargo configuration was also tested with a pedestrian suddenly appearing during navigation; K1 provided a different sensor type from the other platforms.Source: Table IV · p. 6, B. Testing Scenarios and Task… · p. 7
  • CeDR-Nav was evaluated for 80 trials across eight conditions and seven configurations, while DRL-DCLP was evaluated for 40 trials across the four TurtleBot2 configurations.Each condition–method pair was tested for 10 trials.Source: B. Testing Scenarios and Task… · p. 7

C. Real-World Deployment Results

CeDR-Nav completed all 80 real-world navigation trials across seven configurations and three robot platforms without failure. The conclusion highlights reuse of single-source demonstrations and seconds-scale target-specific training.

C. Real-World Deployment Results

  • CeDR-Nav completed all 80 real-world navigation trials across seven configurations and three robot platforms without failure.Go2-C2 also completed all 10 dynamic trials involving a pedestrian entering its path.Source: C. Real-World Deployment Results · p. 8
  • CeDR-Nav navigated narrow passages and densely placed obstacles without collision, while TurtleBot2 under DRL-DCLP frequently collided or timed out in the same scenarios.CeDR-Nav achieved a 100% success rate for both goals across all four TurtleBot2 configurations.Source: C. Real-World Deployment Results · pp. 7–8

VII. CONCLUSION

  • CeDR-Nav derives dedicated target policies from single-source-embodiment demonstrations without additional demonstration collection or online RL training, with target-specific training taking seconds.The conclusion identifies extending the framework to more sensor modalities and navigation scenarios as future work.Source: §VII · p. 8

Related papers

Loading 2610.13007v1…