Taming an End-to-End Autonomous Driving Policy for Urban Navigation of Quadruped Robots

Summary

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.

Abstract Summary

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.

Key Points

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.

METHODS

We formulate urban quadrupedal navigation as goal-conditioned local trajectory planning. Given four surrounding fisheye images, the robot’s proprioceptive state, a high-level directional command, and a two-dimensional goal in the current ego frame, Go2-DrivoR generates $K$ candidate trajectories and selects the one with the highest predicted score.

The output consists of eight $(x,y,\theta)$ waypoints over a four-second horizon, expressed in the current ego frame.

METHODS — Taming an End-to-End Autonomous Driving Policy for Urban Navigation of Quadruped Robots — Fig. 2: Open-loop results on TartanGround. The left panel shows the input fisheye images: front at the top, and left, rear, and right views from left to right along the bottom. The right panel shows candidate trajectories colored by oracle-evaluated PDMS-style scores, the selected trajectory in green , the expert trajectory in dashed blue , and the goal marked by a star . Yellow denotes walkable ground, and pink denotes non-traversable areas. The red rectangle indicates the robot footprint. The “Selected” and “GT” columns report oracle-evaluated sub-scores and aggregate scores for the selected and expert trajectories, respectively. For the selected trajectory, the scoring decoder predicts NOC $=0.993$, DAC $=1.000$, TTC $=0.994$, GEP $=0.924$, and comfort $=1.000$, yielding an aggregate PDMS-style score of $0.959$.

Experiments

The training data consist of 14,192 local planning samples from the TartanGround dataset [ 8 ] . We train the model for 20 epochs using AdamW [ 10 ] on three NVIDIA H200 GPUs with a global batch size of 48. The learning rate is set to $5\times 10^{-4}$, with a two-epoch warm-up phase, followed by cosine annealing.

Open-loop inference is performed on a single NVIDIA H200 GPU, with measured batch latencies of 10.8 ms at batch size 1 and 94.1 ms at batch size 16.

Experiments — Taming an End-to-End Autonomous Driving Policy for Urban Navigation of Quadruped Robots — Fig. 3: Qualitative results on real-world open-loop evaluation. The expert trajectory, selected trajectory, and constant-velocity prediction are shown in dashed blue , green , and dashed orange , respectively. Purple stars mark the four-second lookahead goals. For the selected trajectory, the scoring decoder predicts NOC $=0.993$, DAC $=0.754$, TTC $=0.916$, GEP $=0.906$, and comfort $=1.0$, with an aggregate PDMS-style score of $0.693$.

Sources and Demonstrations

Figures are reproduced from Kim et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.