This is an extremely comprehensive fundamental report, likely the first in the industry to analyze China Robotaxi Players in such detail. It contains extensive technical information, development progress, regulatory updates, operational insights, numerous charts and graphs, as well as UE financial models, all gathered from firsthand interviews.
We are highly confident that no third-party report available in the market matches the depth and quality of analysis presented here.
Due to the report's extensive and detailed nature, we've decided to place the Summary in front of the paywall to help readers understand what content the report covers. Behind the paywall, you'll find all the detailed information corresponding to each summary point.
Executive Summary
Regulations are gradually being relaxed, and several key cities are moving more quickly toward L4-ready rules
Local governments (e.g., Beijing, Shanghai, Guangzhou, Wuhan) have initiated zone-based regulatory pilots, especially for L3+ operations and Robotaxi commercial pilots.
National-level legislation, such as the amendment of the Road Traffic Safety Law, is underway to define AV-specific testing, access, and liability standards—though nationwide policy rollout remains slower.
Representative Initiatives:
Beijing: Demonstration zone phase 4 underway, aiming to cover 1,000+ km².
Guangzhou: 3,240 km² demonstration zone; expected full-city eligibility for Robotaxi by end-2025.
Shanghai: Expanding test areas in Zhangjiang, Huamu, and Expo totalling 758.6 km² in mid-2025.
Currently, over 10 cities allow Robotaxi operations (some fully automated (L4 or above)), but often within fragmented areas. Full-area commercial readiness depends heavily on city-level regulation alignment.
Technology Evolution: Converging on World Models
Training Methodology
Pony.ai: Pioneer in simulation-first learning since 2020. Uses reinforcement learning with its “PonyWorld” cloud model to generate over 10 billion miles of test data weekly. Key belief: imitation learning alone cannot address corner cases or match human cognition.
Other OEMs (Li Auto, XPeng): Initially emphasized real-world data, but are now incorporating cloud-based simulation and world model architectures:
Li Auto mindVLA: Uses hybrid learning from real-world variability to power end-to-end autonomous behavior.
XPeng V6: Progressing toward L3 via reinforcement learning; rollout expected in late 2025.
This industry-wide convergence toward simulation and world model aims to better address long-tail scenarios that cannot be safely or practically collected from real-world data.
Architecture & Backend Autonomy
Pony.ai: Combining multi-modal in-car foundation models with a remote assistant (RA) system to support decisions from the cloud side.
Other OEMs: Increasingly shifting to full-stack integration, including transformer-style models and LLM-based planning logic.
Competitive Landscape: Pony.ai vs. Other OEMs
Pony.ai: Leading China’s L4 Push Through Scale & Simulation
Operational Scale:
The current fleet consists of approximately 270 vehicles, with around half deployed in commercial operations in Guangzhou and Beijing.
By mid-2025, mass production is scheduled to begin with an initial rollout of around 1,000 vehicles. The target is to reach 600 units by the end of 2025 and scale up to 50,000 vehicles within the next 3 to 5 years.
Unit Economics:
Mass production cost expected to fall below $27,778 (RMB 200,000) down from $41,667 (RMB 300,000).
With a backend human intervention ratio of 1:20, the vehicle is estimated to achieve a Miles Per Intervention (MPI) of approximately 500 kilometers.
Break-even expected per vehicle by end-2025.
Business Model Evolution:
Phase 1: Vertically integrated fleet (high CAPEX, high control).
Phase 2: Asset-light operations with third-party fleets and platform revenue.
Phase 3: Pure virtual driver model (e.g., supplying to Didi, AutoNavi).
Financial Forecast (by 2035, assuming 25% market share and 250,000 vehicles):
Revenue: $5.3 billion (RMB 38.2 billion; asset-light)
Net Profit: $1.04-1.25 billion (RMB 7.5–9.0 billion)
Major uncertainty: pace of regulatory approvals.
Other OEMs: Accelerating Into Toward L3 (Conditional Automation)
Current Capabilities:
MPI in the 20–30 km range; target 100–200 km by 2025 year-end.
Li Auto, and XPeng to roll out upgraded platforms (ADS 4.0, mindVLA, V6) by Q3 2025, approaching FSD V13 performance.
Cumulative Training Compute Power by End-2024:
Li Auto: 8.1E FLOPS
XPeng: 2.5E FLOPS (targeting 10E by 2025)
Key risk factors to watch out for
Policy implementation and license issuance remain uncertain in terms of timing and scale, and may face delays; continued monitoring is required.
Future advancement of L2 system developers toward L4 autonomous capability could intensify competition for companies like Pony.ai. There is also potential risk from ride-hailing platforms and other new entrants joining the market.
Underperformance in technical stability may hinder adoption rates and limit penetration.
This research is primarily based on expert interviews across sectors including algorithm development, vehicle deployment, R&D, and project management. However, the research team has not conducted large-scale empirical validation, and some localized information may contain discrepancies. Continuous tracking and verification will be carried out.
Recent stock price surges of relevant companies may carry downside risk.
Note: The RMB to USD exchange rate used in this document is based on the closing rate as of May 22, 2025: 1 USD = 7.2 RMB.
Behind the paywall, you'll find all the detailed information corresponding to each summary point.
