ReaDy-Go: Real-to-Sim Visual Navigation

• ReaDy-Go is a real-to-sim simulation pipeline integrating static 3D Gaussian Splatting with dynamic human GS avatars for robust RGB-based goal navigation.
• It employs photorealistic dynamic dataset generation and a Hybrid A* planner to safely navigate environments populated with moving obstacles.
• The system uses imitation learning on expert supervision to train lightweight, environment-specific policies that effectively bridge the sim-to-real gap.

ReaDy-Go is a real-to-sim simulation pipeline for environment-specific visual navigation with moving obstacles, introduced for RGB-only goal navigation in target deployment environments such as a lobby, library, household, restaurant, or factory (Yoo et al., 12 Feb 2026). It combines a reconstructed static 3D Gaussian Splatting (GS) scene with dynamic human GS obstacles, generates photorealistic dynamic navigation datasets, and trains a goal-conditioned policy intended to be robust both to the sim-to-real gap and to moving humans. The system is organized into three components: a dynamic GS simulator, navigation dataset generation for dynamic environments, and policy learning (Yoo et al., 12 Feb 2026).

1. Problem setting and design objective

ReaDy-Go addresses the setting in which a mobile robot must move from a sampled start to a goal while avoiding both static obstacles and moving humans. The motivating difficulty is twofold. First, visual navigation policies trained in simulation often degrade under sim-to-real transfer because RGB-only control is sensitive to texture, lighting, clutter, geometry quality, and camera appearance. Second, navigation in dynamic environments requires the robot to react safely to moving obstacles rather than only to static geometry (Yoo et al., 12 Feb 2026).

The work is positioned against earlier real-to-sim navigation pipelines based on 3D Gaussian Splatting that assume static scenes or use unrealistic dynamic obstacles. ReaDy-Go therefore targets a more specific regime: target-environment training for dynamic scenes, where the environment is reconstructed from real imagery and dynamic humans are synthesized afterward. The paper’s central operational claim is that photorealistic dynamic simulation can support policies that transfer more effectively to the real target site, particularly when moving obstacles are present (Yoo et al., 12 Feb 2026).

A common misunderstanding is to treat ReaDy-Go as a generic navigation benchmark or as a general-purpose simulator independent of deployment site. The paper instead emphasizes site-specific training: a short monocular video of the target environment is used to build the simulator, and the resulting policy is specialized to that environment. This suggests that ReaDy-Go is best understood as an environment-specific real-to-sim pipeline rather than as a universal navigation model.

2. Dynamic Gaussian Splatting simulator

The simulator begins from a single monocular video of a static deployment environment. Camera poses are estimated with COLMAP in metric scale, with an ArUco marker in the initial frames used to recover metric scale. The static scene is reconstructed as a GS scene, and the implementation uses PGSR rather than vanilla GS because cleaner geometry is needed for occupancy extraction and planning (Yoo et al., 12 Feb 2026).

Dynamic obstacles are represented as animatable human GS avatars. These are obtained from NeuMan videos using HUGS, which disentangles a dynamic human and a static background and parameterizes the human in a canonical space initialized with the SMPL body model. The representation includes human Gaussian parameters, triplane features, learned fully connected networks that estimate Gaussian attributes, and linear blend skinning weights. Given SMPL pose parameters, the human GS can be animated under novel poses (Yoo et al., 12 Feb 2026).

Human motion is synthesized from a desired 2D trajectory using PriorMDM. The trajectory is converted into body root linear velocities and body root rotational velocities, normalized to the HumanML3D representation expected by PriorMDM, and passed through PriorMDM to produce predicted 3D body joint positions. These joints are then fitted to SMPL parameters using SMPLify and transformed into the world frame of the target scene. The result is a full-body animated GS human rather than a rigid moving asset (Yoo et al., 12 Feb 2026).

A key architectural property is representation consistency: both the environment and the dynamic humans are modeled in GS. The paper contrasts this with pipelines that extract meshes and insert non-GS assets through a physics engine. ReaDy-Go instead composes a static scene GS with time-varying human GS avatars and renders the combined scene directly, which the paper presents as a practical route to photorealistic dynamic scenario generation (Yoo et al., 12 Feb 2026).

3. Dynamic dataset generation and planning stack

To generate supervision, ReaDy-Go converts the static GS reconstruction into occupancy maps. Following Splat-Nav, a voxel is marked occupied if it contains all or part of the $1\sigma$ Gaussian ellipsoid. Because reconstructed GS scenes may contain spurious low-confidence Gaussians near the floor, the pipeline adds opacity filtering: opacity is accumulated per voxel, and voxels whose accumulated opacity does not exceed a threshold are treated as free space. The filtered voxel map is then projected into two 2D maps, a robot navigable map and a human walkable map, and both are inflated by a safety margin determined by robot radius or human radius (Yoo et al., 12 Feb 2026).

The human planner samples human start and goal positions so that the resulting human path either crosses or runs parallel to the line connecting the robot’s start and goal. A seed path is computed on the human walkable map using A$^*$, then smoothed with spline-based optimization. That 2D trajectory is passed to the animation module, which produces walking or running motion in SMPL space (Yoo et al., 12 Feb 2026).

The robot expert planner is based on Hybrid A$^*$ and is designed for a unicycle or differential-drive ground robot. Dynamic human occupancy is inserted into the robot navigable map when a human appears in the robot camera’s field of view and comes close to the robot trajectory within a safety margin. Before projection, human GS primitives are cleaned by discarding ellipsoids whose covariance trace exceeds the 75th percentile and then downsampling the remaining primitives by voxel-grid downsampling on their mean positions. The resulting occupancy is inflated using a safety buffer derived from constant-velocity prediction over a 2 s horizon (Yoo et al., 12 Feb 2026).

During Hybrid A$^*$ expansion, the planner uses motion primitives formed from three velocity scales, $\{1/3, 2/3, 1\}$ times maximum velocity, and three steering commands, left, straight, and right. Replanning is reactive: it is triggered either when a predefined replanning period elapses or when the minimum distance between the dynamic obstacle and the look-ahead segment of the current trajectory falls below a safety margin. For each generated scenario, the recorded supervision tuple consists of RGB observation, expert action $(v,w)$, and relative goal position (Yoo et al., 12 Feb 2026).

4. Policy architecture and learning procedure

ReaDy-Go trains an imitation learning policy rather than a reinforcement learning policy. The paper explicitly notes that reinforcement learning appears only as a future direction, not as part of the method itself. At each timestep, the policy receives three consecutive RGB frames, the previous action $(v_{t-1}, w_{t-1})$, and the relative goal position in the robot local frame, $(\Delta x_{t,r}, \Delta y_{t,r})$. The interval between consecutive input images is 0.5 s (Yoo et al., 12 Feb 2026).

The visual encoder has 10 convolutional layers with residual connections and produces a 20-dimensional latent vector per RGB frame. Since the input contains three frames, the visual stream yields $3 \times 20 = 60$ latent dimensions. These are concatenated with the previous action and the relative goal position to form a 64-dimensional vector, which is processed by a 3-layer MLP. The network outputs the current linear and angular velocity, $(v_t, w_t)$ (Yoo et al., 12 Feb 2026).

Training is supervised with mean squared error on expert actions. Optimization uses Adam with learning rate $^*$0. Input images are resized to $^*$1, and the resulting policy has about 2M parameters. The paper interprets the 3-frame input as a way to encode both ego-motion history and motion cues from dynamic obstacles; this suggests that the architecture is deliberately lightweight so that the empirical effect can be attributed primarily to the simulation and data-generation pipeline rather than to a large policy backbone (Yoo et al., 12 Feb 2026).

5. Experimental evaluation and sim-to-real performance

The experiments use three target deployment environments: Outside, Lobby, and Library. For each environment, a monocular video of about 6 minutes was recorded, corresponding to about 1,000–1,500 images, and reconstructed into a GS scene. Training uses 400 episodes per environment, yielding approximately 80k–120k samples per environment, with 50 validation episodes used for checkpoint selection. Two tasks are evaluated: Static, with only static obstacles, and Dynamic, with one or two moving human obstacles. Each task uses 100 simulation episodes and 10 real-world episodes per environment. A run is successful if the robot reaches within 1 m of the goal before a maximum duration of 50 s (Yoo et al., 12 Feb 2026).

In simulation, ReaDy-Go achieved 90% success rate (SR) and 13.43 s average reaching time (ART) on Outside-Static, 98% and 8.93 s on Lobby-Static, and 86% and 11.08 s on Library-Static. On the Dynamic task, the corresponding results were 78% and 18.68 s for Outside, 78% and 17.59 s for Lobby, and 80% and 13.98 s for Library. Vid2Sim was close on several Static settings but degraded more strongly on Dynamic, with 57% SR and 26.88 s ART on Outside-Dynamic, 70% and 21.06 s on Lobby-Dynamic, and 68% and 19.92 s on Library-Dynamic (Yoo et al., 12 Feb 2026).

The real-world results are stronger. ReaDy-Go reached 100% SR and 17.45 s ART on Outside-Static, 90% and 18.93 s on Lobby-Static, and 100% and 11.12 s on Library-Static. On the Dynamic task, it achieved 90% and 20.68 s on Outside, 70% and 26.49 s on Lobby, and 80% and 19.57 s on Library. Vid2Sim, evaluated with the same policy architecture, obtained 60% SR and 28.98 s ART on Outside-Dynamic, 40% and 36.50 s on Lobby-Dynamic, and 60% and 28.54 s on Library-Dynamic. The general navigation baselines GNM, ViNT, and NoMaD typically had much lower SR and higher ART, despite having 9M–30M parameters versus ReaDy-Go’s about 2M (Yoo et al., 12 Feb 2026).

The robot platform in the real experiments was a differential wheeled robot modeled as a unicycle, equipped with a forward-facing fixed ZED2 camera and on-board inference on NVIDIA Jetson Orin NX. Wheel odometry was used to estimate the relative goal position. In zero-shot deployment to an unseen real environment, a policy trained on the combined 1,200 episodes from the three training environments achieved 70% SR and 30.60 s ART on Static, and 50% SR and 35.45 s ART on Dynamic. The authors present this as evidence of generalization potential, while also noting that the longer reaching times indicate the continuing value of scene-specific training (Yoo et al., 12 Feb 2026).

6. Scope, limitations, and relation to similarly named systems

Several limitations are explicit or readily identifiable in the method description. The input environment video is assumed to be static, with dynamics inserted afterward; the dynamic obstacles are humans rather than arbitrary moving objects; the expert planner uses constant-velocity prediction over a 2 s horizon; and the deployed policy assumes access to the relative goal position from odometry rather than operating as a pure image-goal system. The main experiments cover three training environments, and the paper does not provide a formal ablation table. Its strongest ablation-like comparison is against Vid2Sim, where the policy architecture is held fixed and the difference lies in the data-generation pipeline (Yoo et al., 12 Feb 2026).

These limitations matter for interpretation. ReaDy-Go does not claim direct reconstruction of naturally dynamic scenes, and it does not claim physically grounded human-scene interaction through a full physics engine. The system is instead a practical real-to-sim pipeline centered on photorealistic rendering, occupancy extraction, planner-generated supervision, and imitation learning. A plausible implication is that its main contribution lies in dynamic data realism and environment specificity rather than in model-scale innovation.

The name also invites confusion with unrelated systems. ReaDy-Go should not be conflated with “Ready, Bid, Go! On-Demand Delivery Using Fleets of Drones with Unknown, Heterogeneous Energy Storage Constraints” (Talamali et al., 11 Apr 2025), which concerns decentralized auction-based UAV delivery, or with “ReRe: A Lightweight Real-time Ready-to-Go Anomaly Detection Approach for Time Series” (Lee et al., 2020), which addresses streaming anomaly detection. It is likewise distinct from “OpenGo: An OpenClaw-Based Robotic Dog with Real-Time Skill Switching” (Li et al., 2 Apr 2026), which does not mention ReaDy-Go and concerns skill switching for a quadruped robot. Within robotics and real-to-sim navigation, ReaDy-Go specifically denotes the GS-based pipeline for environment-specific visual navigation with moving human obstacles (Yoo et al., 12 Feb 2026).