EIDA: Execution-Interface Dynamics Adaptation for Real-to-Sim-to-Real Robot Navigation
EIDA is a preprint that learns execution-interface dynamics models from real robot velocity-command, pose, and feedback data and embeds them into a lightweight GPU simulator for navigation policy training. On a physical Unitree Go2 it reached the goal without collision in 20/20 static-scene trials versus 4/20 for a standard simulator-trained baseline, and it led learned policies on 100 BARN environments with 89.33% success under global guidance versus 80.78% for Nominal. It offers a cheaper, platform-specific sim-to-real transfer route without detailed actuator modeling.
Can navigation policies trained in a lightweight simulator transfer to a physical robot without reconstructing detailed low-level actuator dynamics?
Velocity-command interfaces hide platform-specific low-level control, so identical commands can produce different trajectories and velocity feedback in simulation versus real hardware, causing drift, turning overshoot, and collisions. Existing sim-to-real methods often focus on visual perception or randomize physics parameters rather than fitting the specific closed-loop execution response of the target platform, leaving action-model mismatch unresolved.
Earlier work used dynamics randomization or SimOpt to sample or update simulation parameter distributions, Grounded Action Transformation or Neural-Augmented Simulation to alter actions or learn trajectory discrepancies, simplified motion models fitted from real data such as Truong et al. and Bono et al., and recurrent world models such as RWM-U for imagined rollouts. These either do not replace simulator transitions with the fitted target-specific response, model full observation transitions expensively, or use learned dynamics for online planning rather than changing the training interface.
EIDA separates execution-interface modeling into two fitted components: an MLP predicts body-frame pose increments from current velocity estimate, current command, and preceding commands plus interval duration; an autoregressive ARX model predicts the next velocity feedback observed by the policy. The pose increments update simulator geometry, while the ARX output feeds policy observations; the policy also receives a short history of velocity estimates. Models are fitted per platform from combined steady-state and policy-driven data, then used to train SAC policies in the FlashNav GPU simulator. At deployment, live sensing and the existing low-level controller replace the learned interface.
On held-out dynamics validation, fitted models reduced endpoint pose error versus the simulator's predefined motion model across 0.1-5 s horizons; at 5 s, position and yaw RMSE dropped 71.3% and 41.3% for Jackal, and 34.4% and 34.9% for Go2. In 100 BARN environments in Gazebo with 300 trials per policy and planner setting, EIDA achieved 81.33% success without global guidance and 89.33% with global guidance, versus Nominal at 71.22% and 80.78%; DWA with global guidance reached 46.00%. Scores were 0.3984 without and 0.4241 with guidance for EIDA. Ablations showed removing the fitted motion model, velocity history, or ARX feedback reduced success by 9.34, 8.34, and 5.67 percentage points without guidance and 16.00, 10.67, and 8.00 with guidance; pose-derived feedback collapsed success by 28.67-45.33 points. On a physical Go2 across two static scenes totaling 20 EIDA trials, EIDA had 0 collisions and 20 goal-reaching successes versus 4/20 baseline successes; among successful runs, path length fell 7.2-11.9% and execution time 20.2-25.4%.
The authors note future work should address long-horizon prediction and uncertainty-aware modeling; the current models are deterministic and do not model localization error or predictive uncertainty. Physical validation is limited to one Unitree Go2 platform, two indoor static scenes, and 20 EIDA trials plus one dynamic pedestrian demonstration, with manual outcome verification; Jackal dynamics results appear simulation-based. Data collection requires per-platform labeled execution data, and no code release or broader multi-robot, outdoor, or longer-horizon deployment is reported.
Mobile-robot OEMs and autonomy teams, especially makers of AMRs, quadrupeds such as Unitree, and Jackal-class research/industrial wheeled platforms, could use EIDA to fine-tune navigation policies to a specific hardware and controller stack without modeling actuators. The method works with existing GPU simulators like FlashNav and existing low-level controllers, so advanced labs could apply it now; production adoption in 1-3 years would require broader validation across fleets, environments, dynamic obstacles, and long-horizon or uncertainty-aware refinements.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.