World Model × Reinforcement Learning: Horizon HSD V2.0 Rebuilds the End-to-End Upper Bound

On June 29, Horizon officially launched the HSD V2.0 version of its full‑scene advanced driver‑assistance system. As the largest OTA upgrade since HSD’s debut, V2.0 introduces six major evolutionary dimensions, 18 new features, and 25 experience enhancements, showcasing a new paradigm of continuous end-to-end system evolution with core capabilities that enable “point‑to‑point navigation and all‑dimensional collision avoidance.” Horizon’s HSD V2.0 is set to roll out to automakers, with the iCAR V27 among the first models to feature it; starting June 30, owners will receive the OTA update in phased releases. The rollout schedules and specific feature experiences for additional partner vehicles will be announced by the respective automakers.

This HSD V2.0 upgrade is not merely a simple accumulation of features; it represents a model‑level evolution built on a world‑model architecture combined with end-to-end reinforcement learning. It delivers all‑round enhancements to the intelligent driving experience, parking and navigation capabilities, comprehensive hazard avoidance, and personalized human–machine interaction, achieving a holistic leap in underlying driving performance. Specifically, in terms of the intelligent driving experience, the vehicle handles dynamic driving scenarios with effortless precision, providing passengers with a smooth, virtually imperceptible ride while significantly improving both on‑road and parking metrics. On the road‑driving front, HSD V2.0 seamlessly integrates end‑to‑end intelligent driving capabilities, excelling in everyday high‑frequency commutes and confidently tackling complex traffic conditions. For parking, the system now ensures “parking available whenever you need it,” handling a wide variety of challenging parking spaces with greater ease and bolstering its obstacle‑avoidance prowess. From a safety standpoint, HSD V2.0 establishes proactive hazard‑avoidance capabilities that cover all targets, directions, speeds, weather conditions, and driving scenarios, substantially expanding the vehicle’s safety perimeter. In addition, this update introduces an AI‑powered, customizable SR skin for the in‑vehicle infotainment system, delivering a fresh, immersive visual experience for intelligent driving.
On a deeper level, as an end-to-end benchmark in China, Horizon HSD has consistently strived to adopt an end-to-end architecture that delivers seamless, all‑scenario capabilities. Version 2.0 leverages an enhanced toolchain to continuously boost the computational efficiency of the Journey 6P chip, further refining its dual‑engine architecture powered by a world model and end-to-end reinforcement learning. This results in a highly consistent, human‑like, and broadly generalizable experience across all scenarios.

Moreover, leveraging an upgraded end-to-end AEB risk‑network model architecture, HSD V2.0 delivers equally significant advances in the realm of active safety. In V2.0, the Occupancy Network (OCC)—originally designed to support NOA navigation‑assisted driving—has been comprehensively extended to cover a broader range of active‑safety use cases, including AEB (Automatic Emergency Braking), AES (Automatic Emergency Steering), and AMAP (Accelerator Pedal Misapplication Prevention). These functions share a unified base model and are continuously refined through high‑precision synthetic scenario data, steadily enhancing the model’s ability to generalize across long‑tail edge cases. As a result, the active‑safety system no longer relies on traditional “whitelist”‑based detection logic, enabling it to handle a wider variety of irregularly shaped obstacle scenarios more effectively and fundamentally improving its risk‑perception and decision‑making capabilities.
If the end-to-end 1.0 version broke the fragmentation inherent in traditional modular intelligent driving systems by leveraging a single large-scale model to deliver universal, all‑scenario driving capabilities and thereby ushered in the first disruptive overhaul of the intelligent‑driving architecture, then HSD V2.0 builds on this foundation by deeply integrating a dual‑engine architecture—combining a world model with end-to-end reinforcement learning—significantly enhancing the system’s capacity for deep reasoning about the driving environment and dynamic prediction, thus achieving a second, profound evolution of end-to-end technology. This upgrade marks a quantum leap in the model’s understanding of physical space, dynamic maneuvers, and risks across the entire domain, elevating intelligent driving from “passively responding to scenarios” to “proactively anticipating the world.” With the industry’s end-to-end race now entering its second half—where the first half was a contest of architecturally innovative, zero‑to‑one implementations—the second half poses a triple challenge: rethinking the end-to-end paradigm, improving iterative efficiency, and demonstrating robust engineering capabilities. As a result, the ability to continuously evolve has become the core competitive barrier for advanced, end-to-end intelligent driving systems.
Leveraging a core architecture that integrates a world model with end-to-end reinforcement learning, Horizon has firmly mastered the “iterative magic wand” at the heart of intelligent driving evolution, pioneering a forward‑looking iterative closed loop based on a uniquely industry‑specific physical‑world foundation model. Built on expert driver data collected from real‑world roads, this system combines the world model and reinforcement learning to iteratively refine its performance in long‑tail scenarios, endowing the intelligent driving system with the characteristics of an autonomous agent—capable of self‑learning, self‑iteration, and continuous evolution. This enables seamless, all‑scenario capability growth across highway and parking functions, as well as between routine conditions and extreme edge cases. Compared with the industry’s fragmented, feature‑by‑feature iteration approach, the core value of HSD V2.0 lies in establishing a sustainable, self‑evolving technological foundation. By harnessing the world model’s spatiotemporal understanding, it steadily bridges the gap between human and machine driving capabilities, enabling continuous iterative upgrades for mass‑produced L2++ intelligent driving and paving the way for the optimal evolutionary trajectory of autonomous driving.
HSD’s continuous evolution has been thoroughly validated in the market. Among models offering the HSD option, as many as 77% of customers actively choose the HSD‑equipped version, casting a vote of confidence with their wallets. The share of HSD‑enabled vehicles in total mileage has already approached 50%, and for the Falcon 700 model powered by HSD, the iCAR V27, the rate of intelligent driving usage soars to 85.14%, demonstrating that users are confidently ceding control of driving to the system. Meanwhile, thanks to deep, co‑designed software‑hardware optimization, HSD reduces latency—from environmental changes to action response—by 34% compared to the industry’s second‑place leader, underscoring its robust potential in advanced driver assistance.

Behind these achievements lies Horizon’s deep, decade‑long technological expertise. Over the past three years, Horizon has had more than 50 papers accepted at top international conferences, with over 100,000 academic citations and more than 2,000 patent applications worldwide. The company also maintains four collaborative R&D labs—covering general AI, embodied intelligence, intelligent agents, and acoustics—whose combined efforts form a robust foundation for HSD’s continuous evolution. In the era of end-to-end systems, autonomous driving platforms can autonomously perceive and understand the world, with their capabilities limited only by the quality and scale of data and training, thereby breaking free from the constraints of engineer‑defined rules. Building on this comprehensive framework, Horizon will continue to accelerate the evolution of its autonomous driving technology, delivering high‑safety, high‑performance solutions that make truly effective advanced driver‑assistance systems accessible to all.

Reposted from Sina Auto

Email:
info@adtfm.com

Phone:
021-20532053

Address:
2nd Floor, No. 150 Jindian Road, Pudong New Area, Shanghai