Grounding Driving VLA via Inverse Kinematics
Grounding Driving VLA via Inverse Kinematics
This research paper introduces a method to improve driving Vision-Language-Action (VLA) models by addressing blind planning issues through counterfactual obstacle insertion analysis and an inverse kinematics-inspired architecture. The proposed approach uses a next visual state prediction objective and a separate Inverse Kinematics Network to suppress reliance on ego status and textual shortcuts during trajectory decoding. A 0.5B model achieves state-of-the-art performance on NAVSIM-v1, NAVSIM-v2, and nuScenes benchmarks using only cameras, without LiDAR or a BEV encoder, improving NAVSIM PDMS and EPDMS scores by 19.0 and 20.4 points respectively compared to the OpenDriveVLA* baseline.
The authors present a new architecture for Driving Vision-Language-Action (VLA) models that uses inverse kinematics-inspired components and future visual state prediction to address blind planning failures, reporting specific performance gains on NAVSIM and nuScenes benchmarks using only camera inputs.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Dataset details reported by the source
- Modalities
- cameras alone Open source S4
- Embodiments
- without LiDAR or a BEV encoder Open source S4
- paper: not_reported
What remains unknown
Dataset licensing and commercial-use rights are not established here.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ture visual states as input is designed to suppress reliance on ego status and textual shortcuts during trajectory decoding. Our 0.5B model achieves state-of-the-art trajectory planning performance on NAVSIM-v1, NAVSIM-v2, and nuScenes using cameras alone, without LiDAR or a BEV encoder. Against the matched OpenDriveVLA* baseline, it improves NAVSIM PDMS and EPDMS by 19.0 and 20.4 points, respectively. Controlled abl
Open source S4
arXiv:2605.21061v2 Announce Type: replace-cross Abstract: Counterfactual obstacle insertion exposes blind planning in representative Driving VLAs: their trajectories remain largely unchanged despite a vehicle appearing directly ahead. An empirically calibrated bias-variance analysis identifies how ego-predictable training examples admit a shortcut through ego status and commands. We address this failure with an inver
Open source S2
Implications for data suppliers
RoboSignal interpretation and collection questions, not statements of buyer demand.
- Confirm the required data type and collection setting with the buyer; this source does not establish a complete collection specification.
- Check the exact dataset version, license, modalities and task coverage against the intended use.
- Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.
Source：arXiv Robotics — research abstracts · arxiv.org