Source-linked AI summary
Taming an End-to-End Autonomous Driving Policy for Urban Navigation of Quadruped Robots
Joochan Kim, Chanuk Yang, Tackgeun You, Ziran Wang, Hwasup Lim
TL;DR
Quadrupedal urban navigation calls for goal-conditioned local planning, but vehicle-centric methods lack explicit local goals and sidewalk-oriented scoring. Go2-DrivoR adapts DrivoR’s goal conditioning and scoring while retaining its core decoders; it improves planning on unseen simulation maps and transfers zero-shot to real-world Go2 trajectory prediction.
Problem
Quadrupedal urban navigation requires short-horizon local planning without the lane centerlines and fixed route geometries assumed by vehicle-centric methods, while existing DrivoR scoring lacks sidewalk desirability and explicit local-goal conditioning.
Source: §I · p. 1Method
Go2-DrivoR conditions trajectory proposals on a local-frame subgoal and adapts drivable-area compliance and ego progress for quadrupedal navigation while retaining DrivoR’s core decoders.
Source: §I · p. 1, A. Go2-DrivoR · p. 1Results
Go2-DrivoR improves planning performance on unseen simulation maps and transfers zero-shot to real-world Go2 trajectory prediction.
Source: §IV · p. 5, 1) Simulator · p. 5, 2) Real-world · p. 5Takeaways & Limitations
DrivoR’s core decoders can be retained for destination-based local planning with quadrupedal robots.
Source: §IV · p. 5Evaluation relies on expert-derived lookahead goals and open-loop real-world prediction, with no explicit dynamic-agent modeling.
Source: §IV · p. 5
Abstract
from arXiv · showhide
We present Go2-DrivoR, a goal-conditioned adaptation of the end-to-end autonomous driving trajectory planning framework DrivoR for urban navigation with quadrupedal robots. By conditioning trajectory generation on a local-frame subgoal through a goal token and adapting the vehicle-centric scoring formulation, the method extends DrivoR to short-horizon goal-conditioned local planning without redesigning its core decoders. Specifically, we redefine drivable-area compliance for sidewalk-oriented navigation and reformulate the original ego progress term as goal-conditioned ego progress. Trained exclusively on TartanGround simulation data, Go2-DrivoR improves waypoint-conditioned planning performance on unseen simulation environments and transfers zero-shot to open-loop real-world trajectory prediction.
I. INTRODUCTION
Go2-DrivoR adapts DrivoR to quadruped urban navigation by conditioning trajectory generation on a local goal and adapting vehicle-centric scoring while retaining the core decoders. It is evaluated on simulation environments and real-world Go2 logs withheld from training.
I. INTRODUCTION
- Quadruped navigation lacks the lane centerlines and fixed route geometries assumed by vehicle-centric methods, while requiring local planning under legged-motion and tracking constraints.The policy must use visual observations and proprioceptive states while accounting for obstacles, pedestrians, and surface variation.Source: §I · p. 1
A. Go2-DrivoR
- A local-frame waypoint conditions trajectory proposals through a goal token, while sidewalk-oriented DAC and goal-conditioned ego progress adapt scoring without redesigning DrivoR’s core decoders.The method retains robot-footprint-aware safety criteria and aligns trajectory evaluation with the supplied local goal.Source: §I · p. 1, A. Go2-DrivoR · p. 1
I. INTRODUCTION
- Training uses TartanGround simulation data exclusively, with real-world Go2 logs withheld for zero-shot open-loop trajectory evaluation.The evaluation covers seen and unseen simulation environments as well as real-world logs.Source: §I · p. 1
1) Ego-Proprioceptive and Goal Token Extraction:
The model represents a four-second local waypoint as a goal token alongside an ego-state token. Their fusion conditions trajectory queries that feed DrivoR’s unchanged decoder.
1) Ego-Proprioceptive and Goal Token Extraction:
- Each local-planning sample’s two-dimensional goal is the waypoint four seconds ahead, transformed into the current ego frame and normalized by a 4 m distance scale.The goal token is obtained by projecting this normalized subgoal into latent space.Source: 1) Ego-Proprioceptive and Goal Token… · pp. 1–2
- Proprioceptive signals are projected into an ego-state token, which is concatenated with the goal token and projected back to the model dimension.The physical state includes relative ego pose, linear and angular velocities, acceleration, and command inputs.Source: 1) Ego-Proprioceptive and Goal Token… · pp. 1–2
- The fused token is added residually to each trajectory query, and the unchanged decoder produces 64 candidate trajectories of eight future waypoints over four seconds.Each waypoint contains (x,y,θ) and is sampled at 0.5-second intervals.Source: 1) Ego-Proprioceptive and Goal Token… · p. 2
2) Trajectory Scoring for Urban Navigation:
Go2-DrivoR adapts trajectory scoring for goal-conditioned quadrupedal navigation by replacing vehicle-centric criteria with robot-footprint-aware safety checks and sidewalk- and goal-aware measures. Its scorer uses five supervised components and combines predicted sub-scores to rank trajectories.
Trajectory Scoring for Urban Navigation
- The adapted scorer supervises NOC, DAC, TTC, GEP, and comfort, omitting driving-direction compliance because destination-based quadrupedal navigation has no prescribed road or lane direction.NOC and TTC retain DrivoR’s formulations but use the quadrupedal robot footprint for collision and time-to-collision evaluation.Source: 2) Trajectory Scoring for Urban… · p. 2
- DAC scores trajectories by semantic region: 1 for all-sidewalk waypoints, 0.5 for road but no out-of-bounds waypoints, and 0 otherwise.These labels form the DAC oracle targets for scoring supervision.Source: 2) Trajectory Scoring for Urban… · p. 2
- GEP combines original ego progress with goal-directed progress, measured by projecting a candidate endpoint displacement onto the unit direction toward the local goal.The formulation retains the original progress signal while additionally rewarding motion toward the goal; α = 0.5 in experiments unless specified otherwise.Source: 2) Trajectory Scoring for Urban… · p. 2
- In the open-loop TartanGround example, the scoring decoder predicts for the selected trajectory NOC = 0.993, DAC = 1.000, TTC = 0.994, GEP = 0.924, comfort = 1.000, yielding an aggregate PDMS-style score of 0.959.Source: Figure 2 · p. 3
- At inference, predicted sub-scores are combined in a log-domain composite score for proposal ranking; the score itself is not directly optimized, and component heads use binary cross-entropy-with-logits supervision.The loss is applied component-wise against oracle targets and predicted logits.Source: 2) Trajectory Scoring for Urban… · p. 3
B. Simulation and Real-World Data Collection
The planner is trained on high-fidelity simulation data and evaluated for zero-shot transfer on real-world Unitree Go2 logs using standardized inputs.
B. Simulation and Real-World Data Collection
- The planner is trained on high-fidelity simulation data and evaluated for zero-shot transfer on real-world Unitree Go2 logs.Both domains use standardized preprocessing and serialization to unify sensor modalities, coordinate systems, and routing interfaces.Source: B. Simulation and Real-World Data… · p. 4
- A standardized preprocessing and serialization pipeline prepares simulation and real-world data for the planner.The pipeline unifies sensor modalities, coordinate systems, and high-level routing interfaces in the planner’s input format.Source: B. Simulation and Real-World Data… · p. 4
1) TartanGround:
TartanGround data are converted into local planning samples with four fisheye views, transformed kinematic states, and a four-second lookahead goal.
1) TartanGround:
- TartanGround provides photorealistic multi-camera renderings, IMU telemetry, and dense semantic point clouds across synthetic environments.Pinhole images are reprojected into front, left, right, and rear fisheye views, while kinematics are transformed into ENU and FLU frames.Source: 1) TartanGround · p. 4
- A sliding-window procedure converts continuous trajectories into local planning samples conditioned on the expert pose four seconds ahead.The lookahead pose is expressed as a goal in the current ego frame, and the same waypoint’s lateral displacement determines a synthesized four-way command.Source: 1) TartanGround · p. 4
- Semantic point clouds are projected onto a 2D grid to construct boundary, roadway, and walkway vector layers for scoring.Source: 1) TartanGround · p. 4
2) Real-world:
Real-world Go2 logs pair four-camera video with synchronized robot states and operator-driven trajectories for open-loop evaluation.
2) Real-world:
- Four GMSL Owl cameras mounted around the Go2 provide front, left, rear, and right views during human-teleoperated data collection.The recorded robot motion serves as the expert trajectory for open-loop evaluation.Source: 2) Real-world · p. 4
- Visual streams are aligned with proprioceptive logs through temporal cross-referencing, and raw fisheye images are used without rectification.Source: 2) Real-world · p. 4
- Robot states are converted from the native FRD frame to FLU, while goal-oriented commands are inferred from lookahead targets using a lateral threshold.Source: 2) Real-world · p. 4
A. Implementation Details
Go2-DrivoR outperforms the reported baselines on simulation aggregate scores and real-world trajectory metrics, with qualitative examples illustrating goal-directed selection. The evaluation uses simulation training and open-loop testing, including real-world logs withheld from training.
A. Implementation Details
- Training uses 14,192 TartanGround samples for 20 epochs, with open-loop inference measured at 10.8 ms for batch size 1 and 94.1 ms for batch size 16.Training uses AdamW on three NVIDIA H200 GPUs; inference latency is measured on one NVIDIA H200 GPU.Source: A. Implementation Details · p. 4
1) Simulator:
- The simulator evaluation compares Go2-DrivoR with constant velocity and zero-shot iPlanner and reports NOC, TTC, DAC, GEP, comfort, and an adapted PDMS-style aggregate score.Real-world testing uses withheld Go2 logs without additional training and reports displacement and heading metrics.Source: 1) Simulator · pp. 4–5, 2) Real-world · p. 5
- Go2-DrivoR achieves higher aggregate scores than constant-velocity and zero-shot iPlanner on seen and unseen maps.Against iPlanner, it also has higher average NOC, TTC, DAC, and comfort scores on both splits, though iPlanner is slightly higher on unseen-map GEP.Source: 1) Simulator · p. 5
- The simulator example reaches the local goal while remaining in the allowable region, scoring 0.933 despite differing from the logged expert path.In the real-world intersection example, Go2-DrivoR selects a turn toward the goal while constant velocity continues along the initial heading.Source: 1) Simulator · p. 5, 2) Real-world · p. 5
2) Real-world:
- Go2-DrivoR outperforms constant velocity across all reported real-world metrics and both scene categories.The reported metrics are ADE, FDE at 1s, 2s, and 4s, and Heading MAE; the largest reduction is at FDE@4s.Source: 2) Real-world · p. 5
C. Ablation Studies
Go2-DrivoR transfers to unseen simulation maps and zero-shot real-world trajectory prediction, while ablations support explicit goal conditioning and blended progress in the evaluated setting. The conclusion notes evaluation-scope limitations and identifies future work.
C. Ablation Studies
- Removing the goal token substantially degrades trajectory prediction accuracy and goal-oriented metrics, even with blended progress retained.The authors suggest coarse directional commands may leave local paths ambiguous in open urban spaces, but say the source of the larger trajectory errors remains unclear.Source: C. Ablation Studies · p. 5
- Both EP-only and Goal-only perform worse than Go2-DrivoR’s blended progress formulation in the evaluated setting.Goal-only does not match the blended formulation in trajectory accuracy or progress quality.Source: C. Ablation Studies · p. 5
IV. CONCLUSION
- Go2-DrivoR improves planning on unseen simulation maps and transfers zero-shot to real-world Go2 trajectory prediction.It is trained exclusively on TartanGround; evaluation uses expert-derived lookahead goals and open-loop real-world prediction.Source: §IV · p. 5
- The conclusion identifies expert-derived lookahead goals, open-loop real-world evaluation, and no explicit dynamic-agent modeling as limitations.Future work targets closed-loop social navigation, temporal scene modeling, and improved sim-to-real adaptation.Source: §IV · p. 5
APPENDIX
DrivoR represents scenes with camera-aware register tokens, generates candidate trajectories from scene and ego-state information, and uses separate decoders for trajectory generation and scoring.
A. Preliminary: DrivoR
- DrivoR maps multi-camera images and ego states to candidate trajectories and their scores using perception, trajectory, and scoring components.Its architecture includes a perception encoder, trajectory decoder, and disentangled scoring decoder.Source: A. Preliminary: DrivoR · p. 6
1) Visual Feature Extraction:
- Camera-specific learnable registers compress patch-level visual features into a compact scene representation, with each camera processed independently.The representation uses N_c cameras, N_r registers per camera, and feature dimension D.Source: 1) Visual Feature Extraction · p. 6
2) Trajectory Decoder:
- The trajectory decoder uses scene-attending queries to generate K ego-frame trajectories and trains only the candidate closest to the expert trajectory under a winner-takes-all objective.Ego-state information, including pose, velocity, acceleration, and driving command, is added to the trajectory queries.Source: 2) Trajectory Decoder · p. 6
3) Scoring Decoder:
DrivoR separately scores generated trajectories against safety, drivability, direction, progress, collision-time, and comfort criteria. The related-work discussion contrasts this candidate-ranking approach with other robot-navigation methods.
3) Scoring Decoder:
- DrivoR detaches each generated trajectory before a separate scoring decoder processes it, preventing scoring gradients from reaching the trajectory decoder while allowing shared perception learning.The scoring decoder re-embeds each trajectory and cross-attends to the same scene representation.Source: 3) Scoring Decoder · p. 6
- Six PDMS-derived components score candidates: NOC, DAC, DDC, EP, TTC, and comfort; inference combines predicted sub-scores to select the highest-scoring trajectory.DDC is predicted and supervised but has zero aggregation weight in the NAVSIM-v1 PDMS formulation used by DrivoR.Source: 3) Scoring Decoder · p. 6
1) Robot Navigation:
- Robot-navigation alternatives include classical local planners, goal-directed path prediction, cross-robot visual policies, and language-conditioned or foundation-model approaches.This work instead generates and ranks multiple short-horizon goal-conditioned quadrupedal trajectories using disentangled safety, drivability, and goal-progress criteria.Source: 1) Robot Navigation · p. 6
2) End-to-End Autonomous Driving:
End-to-end driving methods plan from multimodal observations, but their scoring is primarily vehicle-centric. The paper situates its non-reactive evaluation alongside interactive platforms and datasets for ground-robot research.
2) End-to-End Autonomous Driving:
- End-to-end driving approaches include visual-LiDAR fusion, diffusion-based trajectory planning, sparse scene representations, and DrivoR’s camera-aware registers with separate generation and scoring decoders.Source: 2) End-to-End Autonomous Driving · p. 6
- Vehicle-oriented drivable-area and ego-progress criteria do not explicitly measure progress toward an arbitrary local destination, so the work adds local-goal conditioning and goal-directed progress for quadrupedal navigation.Source: 2) End-to-End Autonomous Driving · p. 6
3) Evaluation Platforms and Datasets:
- Non-reactive evaluation leaves future observations unaffected by policy outputs, whereas closed-loop platforms feed outputs back into the environment; closed-loop quadrupedal navigation is left for future study.The discussion names NAVSIM for non-reactive logged-scene evaluation and CARLA and MetaUrban as interactive platforms.Source: 3) Evaluation Platforms and Datasets · pp. 6–7
- TartanGround provides multimodal data from diverse photorealistic simulations, which this work processes into short-horizon local-planning samples for training and non-reactive open-loop evaluation.Source: 3) Evaluation Platforms and Datasets · pp. 6–7