PaperlayerView exact revision on arXiv

Source-linked AI summary

EIDA: Execution-Interface Dynamics Adaptation for Real-to-Sim-to-Real Robot Navigation

Yiwei Qian, Shanze Wang, Qingyuan Hu, Xinming Zhang, Wei Zhang

arXiv:2610.01219v1cs.RO

TL;DR

Simulation-trained navigation policies face a transfer gap when execution produces motion and velocity feedback unlike the training model. EIDA fits both responses from target-platform data for lightweight-simulator training, and achieves strong benchmark and physical-trial outcomes without detailed actuator simulation.

  • Problem

    Controller response, latency, and recent commands can make identical velocity commands produce different trajectories and feedback, complicating transfer from simulation to physical robots.

    Source: §I · p. 1

  • Method

    EIDA fits body-frame pose increments and a separate velocity-feedback model from target-system executions, using velocity history during simulation training while retaining the existing controller at deployment.

    Source: §I · pp. 1–2, §IV · p. 2

  • Results

    EIDA achieved the highest success rate and score under both planner settings, and reached the goal without collision in all 20 physical Go2 trials.

    Source: D. Ablation Studies in Simulation · p. 6, §VI · p. 8

  • Takeaways & Limitations

    The framework supports platform-specific adaptation through an existing velocity-command interface without reconstructing low-level actuation dynamics.

    Source: §VI · p. 8

    The models are deterministic and do not explicitly model localization error or predictive uncertainty; future work targets long-horizon prediction and uncertainty-aware modeling.

    Source: A. Execution-Interface Dynamics… · p. 3, §VI · p. 8

Abstract

from arXiv · show

Simulation-to-robot transfer can fail when velocity commands produce motion and feedback that differ from those modeled during policy training. We present execution-interface dynamics adaptation (EIDA), which fits these responses from target-platform execution data without reconstructing actuator dynamics. A model of body-frame pose increments updates simulator geometry, while a separate model predicts the velocity feedback observed by the policy; a short history of velocity feedback is included in the policy input. The fitted models are used within a lightweight GPU-parallel simulator. On the full Jackal and Go2 validation sets, the fitted models reduced position and yaw prediction errors relative to the simulator's predefined motion model. Across 100 benchmark navigation environments evaluated in a separate physics-based simulator, EIDA achieved the highest success rate and navigation score among the compared learned policies, both with and without global guidance. Feedback ablations further supported the need to match policy-facing velocity estimates. On a physical Unitree Go2, EIDA reached the goal without collision in all 20 static-scene trials, compared with 4 of 20 for the baseline. These results show that execution-interface adaptation can improve navigation transfer without detailed actuator simulation.

I. INTRODUCTION

Simulation-trained navigation policies may not transfer because real execution changes both robot motion and the velocity feedback available to the policy. EIDA fits these responses from target-system executions for lightweight-simulator training while retaining the existing controller at deployment.

I. INTRODUCTION

  • Controller response, latency, and recent commands can make identical velocity commands produce different trajectories, with small errors more consequential at higher speeds.These differences can contribute to drift, turning overshoot, or collisions.Source: §I · p. 1
  • The approach targets the transfer gap between lightweight simulation and physical robots without reconstructing detailed low-level actuator dynamics.The paper motivates fitting pose evolution together with velocity feedback available to the policy.Source: §I · p. 1
  • EIDA fits body-frame pose increments to update simulator geometry and separately predicts policy-facing velocity feedback from target-system execution data.Recent velocity observations help the policy account for history-dependent execution responses.Source: §I · pp. 1–2
  • Validation spans wheeled and legged platforms and evaluates navigation transfer in simulation and physical robot trials.The contributions describe efficient parallel policy training and validation across platforms.Source: §I · pp. 1–2

II. RELATED WORK

Prior work addresses dynamics mismatch through simulation variation, learned motion corrections, or planning-time models; EIDA instead focuses on the execution interface used during policy training. The preliminaries define local navigation observations, velocity commands, pose increments, and a reward combining progress with terminal outcomes.

II. RELATED WORK

  • Dynamics randomization and SimOpt vary simulated systems, while other approaches transform actions, correct trajectory discrepancies, or fit navigation motion models,.NeRD replaces low-level dynamics and contact solvers with robot-centric neural dynamics.Source: §II · p. 2
  • Learned forward models support planning-time trajectory and collision-risk prediction,, while Robotic World Model methods optimize policies with imagined rollouts [17], [18].Source: §II · p. 2
  • EIDA updates simulator geometry with fitted body-frame pose increments and separately predicts policy-facing velocity feedback, unlike methods that transform actions or replace broader simulator dynamics.Source: §II · p. 2

III. PRELIMINARIES

  • The local-navigation policy receives LiDAR ranges, target distance and bearing, and estimated velocity, then issues a planar velocity command to a low-level controller.The command is requested motion, while the velocity estimate is feedback available to the policy.Source: §III · p. 2
  • Policy actions map to physical commands, pose changes are represented in the starting body frame, and the reward combines goal progress with success and collision outcomes.Collision takes precedence when both terminal conditions hold; no additional time penalty is applied.Source: §III · p. 2

IV. METHODOLOGY

EIDA defines an execution interface from target-system execution data for policy training in simulation. At deployment, live sensing and the existing controller replace the fitted responses.

IV. METHODOLOGY

  • The execution interface maps policy velocity commands to body-frame motion and velocity feedback with the existing low-level controller and state estimator in the loop.The fitted interface supports simulation training; live sensing and the existing controller are used at deployment.Source: §IV · p. 2, Figure 2 · p. 3

A. Execution-Interface Dynamics Adaptation

EIDA uses recent control and feedback inputs to predict pose increments and a separate ARX model to predict velocity feedback. During recursive simulation, these predictions update physical pose and the next policy input, respectively.

A. Execution-Interface Dynamics Adaptation

  • The pose model uses the current velocity estimate, current command, and preceding commands, then predicts a body-frame translation and yaw increment that updates simulated pose once.Lateral displacement is retained even for platforms without a lateral command.Source: A. Execution-Interface Dynamics… · p. 3
  • A separate causal ARX model predicts the velocity feedback observed by the policy from recorded velocity estimates and commands.Its full coefficient matrices capture cross-axis dependence.Source: A. Execution-Interface Dynamics… · p. 3
  • In recursive simulation, pose prediction advances physical state while predicted velocity feedback becomes the next policy input; both models are deterministic and omit explicit localization error and predictive uncertainty.Ray casting, target geometry, rewards, and collision checks use the simulated physical pose.Source: A. Execution-Interface Dynamics… · p. 3

B. Model Identification

The identification data pair command and velocity histories with adjacent poses, combining designed excitation and navigation-like trajectories. Separate observation and pose fitting, including recursive multistep training, prepares the models for simulator use.

B. Model Identification

  • Manual excitation and policy-shaped trajectories are paired with command history, velocity estimates, adjacent poses, pose increments, and next-step velocity targets.The two trajectory sources capture designed motion responses and navigation-like command changes.Source: B. Model Identification · p. 3
  • Training windows require complete command histories within continuous valid segments, and pose gaps are not interpolated.Related trajectory repeats stay together in data splits; normalization and model parameters use training data only.Source: B. Model Identification · pp. 3–4
  • The observation model is fitted by weighted ridge regression and frozen before pose-model training, which uses a normalized, weighted single-step loss.The loss uses Smooth-L1, with weights balancing data families and trajectory groups.Source: B. Model Identification · p. 4
  • Multistep fitting recursively replaces measured velocity inputs with ARX predictions, then compares composed predicted and reference increments at the endpoint.Single-step and multistep optimizer updates are interleaved after initial single-step training.Source: B. Model Identification · p. 4

C. Policy Learning and Deployment

Policies are trained with fixed execution-fitted models and recent velocity observations, then deployed using live sensing and the robot’s existing low-level controller. Neither fitted model runs online at deployment.

C. Policy Learning and Deployment

  • A soft actor-critic policy is trained with both execution-fitted models fixed.Batched model inference supports parallel environments, and policy gradients are not propagated through the simulator.Source: C. Policy Learning and Deployment · p. 4
  • Recent velocity observations let policy decisions account for motion responses shaped by preceding commands.The policy input includes velocity history, unlike the pose model’s command history, which represents control inputs.Source: C. Policy Learning and Deployment · p. 4
  • The policy action is mapped to a physical command, and the execution-fitted simulator returns the next observation and reward during training.SAC optimizes the entropy-regularized expected discounted return.Source: C. Policy Learning and Deployment · p. 4
  • At deployment, live LiDAR, localized target geometry, and velocity-estimate history replace the simulator observation interface, while the existing low-level controller remains.Live estimator outputs replace predicted feedback, and the history resets at episode boundaries.Source: C. Policy Learning and Deployment · p. 4

V. EXPERIMENTS

Experiments fit platform-specific dynamics from execution data and train policies in a GPU-parallel simulator. The supplied passages describe platform data, model and policy configurations, training time, and the prediction-validation figure.

A. Data Collection and Training

  • Nominal uses the predefined motion model and current velocity feedback, without execution-fitted dynamics or velocity-history input.Jackal and Go2 policies were trained separately using each platform’s execution data in FlashNav, a GPU-parallel two-dimensional simulator.Source: A. Data Collection and Training · p. 4
  • Both platforms’ identification datasets combine steady-state excitation and policy-driven navigation, split by trajectory group; navigation recordings receive 80% of training weight.Go2 has 26,495 training and 19,842 validation samples; Jackal has 32,225 training and 6,401 validation samples.Source: A. Data Collection and Training · p. 4
  • The pose predictor uses a two-layer 256-unit MLP, while an independent ARX feedback model is fitted by weighted ridge regression.The ARX regularization coefficient is 10^-3.Source: A. Data Collection and Training · p. 4
  • Fig. 3 plots mean endpoint RMSE over prediction horizons for Jackal and Go2 validation sets, with one-standard-deviation error bars across three pose-model seeds.Source: Figure 3 · p. 5
  • Two million policy-training transitions took 4.02 ± 0.03 min for Jackal and 4.97 ± 0.01 min for Go2 across three seeds.Both evaluated Go2 policies used a 0.75 × 0.36 m robot footprint.Source: A. Data Collection and Training · p. 5

B. Dynamics Model Validation

Across 100 BARN worlds, EIDA led the compared learned policies in navigation success and score under both planner settings. The execution-fitted model also lowered validation pose errors, though Go2 position gains did not hold consistently within the deployment command envelope.

B. Dynamics Model Validation

  • At 5 s, Jackal position/yaw RMSE fell 71.3%/41.3%, and Go2 fell 34.4%/34.9%; Fig. 3 shows horizon-wise validation errors.The fitted response had lower position and yaw RMSE at all evaluated horizons on both full validation sets.Source: B. Dynamics Model Validation · p. 5

C. Navigation Performance in Simulation

  • Navigation policies were evaluated three times in each of 100 BARN worlds, with success requiring a collision-free goal reach within 1.0 m and 100 s.The comparison included DWA, E-Band, Nominal, and RWM-U; learned policies were tested with and without global guidance.Source: C. Navigation Performance in Simulation · pp. 5–6
  • EIDA led learned policies in success rate and score with and without guidance; versus Nominal, success rose 10.11 points without guidance and 8.55 with it.With guidance, success was 89.33% ± 2.33% and score was 0.4241.Source: C. Navigation Performance in Simulation · p. 6, Figure 5 · p. 6

D. Ablation Studies in Simulation

The BARN ablations isolate motion-model fitting, velocity history, and feedback source within the complete EIDA configuration. Removing each component reduced success and navigation score, while comparisons favored ARX feedback over alternatives.

D. Ablation Studies in Simulation

  • Table II varies motion model, velocity history, and feedback source in a 2 × 2 × 2 ablation; EIDA is the Fitted/Yes/ARX configuration.Configurations were evaluated on 100 BARN worlds with three repeats, with and without global guidance.Source: Table II · p. 6, D. Ablation Studies in Simulation · p. 6
  • Success fell 9.34 and 16.00 percentage points without and with global guidance when motion fitting was removed, 8.34 and 10.67 with history removed, and 5.67 and 8.00 with ARX replaced; all removals also lowered score.Source: D. Ablation Studies in Simulation · p. 6
  • Velocity history supplies recent-response information because similar current commands and estimates can lead to different motion after different preceding execution sequences.The pose predictor conditions on recent commands, and the effects of components were not uniformly additive across partial configurations.Source: D. Ablation Studies in Simulation · p. 6
  • Pose-derived feedback reduced success by 28.67 percentage points without guidance and 45.33 with it; the latter setting's score fell 52.5%, while ARX outperformed both alternatives.Pose-derived feedback averages predicted displacement over one control interval; ARX models the velocity estimate available to the deployed policy.Source: D. Ablation Studies in Simulation · pp. 6–7

E. Real-World Deployment

In two indoor Go2 scenes, EIDA was more successful than the Nominal baseline and had shorter paths and execution times among successful trials. The evaluation also illustrates navigation under a sudden pedestrian obstruction.

E. Real-World Deployment

  • Onboard Go2, the policy issued velocity commands at 10 Hz through the existing low-level interface; EIDA used current plus five past velocity observations, versus current observations for Nominal.The platform used a Livox MID-360 LiDAR and Jetson Orin NX; FAST-LIO2 supplied pose estimates and Sport supplied body-frame velocity estimates.Source: a. Platform and Deployment Setup · p. 7
  • 20/20 versus 4/20 trials: EIDA reached the goal without collision; among successful trials, it reduced mean path length by 7.2–11.9% and execution time by 20.2–25.4% across the two scenes.Source: Table III · p. 7, c. Experimental Results · p. 7
  • The two indoor turf scenes tested a curved passage and separated obstacle clusters; each policy ran ten trials per scene, with a 0.3 m goal radius and 30 s limit.Goal required reaching the target without collision, verified by manual review.Source: b. Scenarios and Evaluation Protocol · p. 7
  • The Nominal trials showed oscillatory motion and obstacle contact, whereas EIDA traversed the curved passage and obstacle-cluster gaps without collision.Source: c. Experimental Results · p. 7
  • In the separate pedestrian-obstruction demonstration, EIDA adjusted motion as free space changed, maintained clearance, and subsequently reached the goal.Source: c. Experimental Results · p. 7

VI. CONCLUSION

The conclusion presents EIDA as a way to transfer navigation policies through lightweight simulation using target-system execution responses, without reconstructing low-level actuation dynamics. Validation included model, benchmark, and physical Go2 experiments.

VI. CONCLUSION

  • EIDA reached the goal without collision in all 20 physical Go2 trials, with shorter paths and lower execution times than Nominal among successful trials; future work targets long-horizon prediction and uncertainty-aware modeling.Source: §VI · p. 8
Loading 2610.01219v1…