NIO World Model New Version Pushed Today: 700,000 Owners Upgrade Simultaneously, End-to-End Kinematic Modeling on Board

On June 18, NIO officially began rolling out the new version of its World Model (NWM). This update covers more than 700,000 users, going all the way back to Banyan platform owners who bought their cars in 2022—which means people who purchased a car four years ago can now enjoy an intelligent driving experience almost identical to that of the 2026 ES9.

This alone is worth highlighting. At a time when most emerging automakers treat new features as exclusive selling points for new cars, NIO’s “no platform distinction, upgrade together” approach essentially uses a unified model architecture to smooth out hardware differences. Of course, the prerequisite is that the underlying NX9031 hardware reserved sufficient capacity.

Three-Layer Training Framework: Filling the Gap with Supervised Fine-Tuning

The previous version of NWM centered on “World Model + Closed-Loop Reinforcement Learning”. The World Model handles long-term reasoning and can simulate 216 possible trajectories in 100ms; Closed-Loop Reinforcement Learning iterates in simulated environments, teaching the model basic driving common sense.

But after running this architecture for over half a year, the NIO team probably noticed an old problem: Reinforcement Learning can make the model “avoid accidents,” but doesn’t necessarily make it “drive like a human.” No matter how meticulously you craft RL’s reward function, it’s still fundamentally metric optimization, leading to strategies that are “compliant but stiff”—brake points are precise, but braking is uncomfortable; lane changes are timed correctly, but the maneuvers lack the smoothness of an experienced driver.

The added middle layer in this update is Supervised Fine-Tuning (SFT). The three-layer structure now looks like this:

• World Model: Understands the physical world, predicts the future, and provides generative scene representations
• Supervised Fine-Tuning: Uses high-quality human driving behavior data to finely sculpt behavior
• Closed-Loop Reinforcement Learning: Ensures a safety floor and compliance via simulated runs

For those familiar with LLM training processes, this combo will sound familiar—essentially an autonomous driving translation of Pre-train → SFT → RLHF. The advantage of SFT here is that it feeds “human-like behavior” directly into the model as an imitation target, with no detours. RL excels at getting from 0 to 80 points, while SFT polishes that machine-like “80” into a human-like “90”.

According to NIO, the three-layer framework gives the model “a high floor, human likeness, and compliance” simultaneously. In simpler terms: RL ensures the floor, SFT adds human-like qualities, and the World Model handles compliance and generalization. Each plays its role.

Direct Steering and Pedal Output: Removing the Intermediate Trajectory Layer

The most hardcore technical element of this release is China’s first intelligent driving system to output steering wheel angles and acceleration/brake pedal signals directly.

To appreciate its significance, you first have to understand the traditional intelligent driving chain:

Sensor input → Perception → Prediction → Planning (output trajectory) → Control (trajectory → steering + pedals) → Vehicle execution

Even in so-called “end-to-end” approaches, most simply merge perception, prediction, and planning into one model, but still output a sampled trajectory, which then goes to a separate control module to be translated into steering and pedal inputs.

The problem? A trajectory is an “intermediate representation.” It assumes the vehicle can follow it precisely, but in reality, lateral dynamics, tire grip, slopes, wind resistance—all are variables. Even the best controllers introduce tracking errors. Worse, the planning layer doesn’t know exactly how precisely the control layer can follow, so the “optimal trajectory” often gets compromised in execution.

NIO’s approach is to remove the trajectory layer entirely, letting the model go from sensor data straight to steering wheel angle and pedal signals—this is end-to-end kinematic modeling in the truest sense.

The benefits are threefold:

1. Shorter path: One fewer intermediate representation means less information loss
2. Lower latency: Minimizes conversion between inference and translation
3. Finer vehicle control: The model directly observes “how humans steer and brake” during training, learning the full motion control loop end-to-end

This mirrors the leap in machine translation from “phrase-based statistical” to “neural end-to-end”: the more intermediate representations, the more error accumulates; letting the model learn the complete mapping yields better results.

Of course, the trade-offs are obvious—reduced interpretability and increased debugging difficulty. Failure modes are no longer “planning misjudged” or “controller overshot” but “some gradient in the black box went in the wrong direction.” NIO’s confidence in pushing this likely comes from the robustness of its closed-loop simulation infrastructure (NSim) and data pipeline.

Overhead Sign Recognition: Eliminating the Last Piece of High-Precision Map Dependency

Scenarios like tidal lanes and variable lanes were once among the last strongholds of high-precision maps—rules are dynamic, so pre-annotated maps can’t handle them; you must recognize the overhead electronic sign in real-time to know the current instructions.

NIO claims to be the first automaker to achieve real-time recognition of tidal/variable lane overhead signs using only self-developed solutions. Without relying on high-precision maps or enhanced navigation, the system extracts this semantic information purely through cameras and models.

The significance? It further reduces the survival space of “map-provider solutions.” Once a model can understand the real-time meaning of signs directly from raw pixels, the “lane-level priors” maps provide are relegated to being mere redundancy checks.

Human-Likeness vs. False Braking: A Contradictory Pair of Metrics

NIO also put forth an interesting claim: achieving “industry-leading” performance in both false braking and risk intervention rates—two inherently contradictory metrics.

Anyone who has worked on ADAS understands the conflict. To reduce risk interventions (false negatives), the system must treat more suspicious objects as genuine threats, which increases false braking (false positives). Conversely, lowering false braking requires raising activation thresholds, which naturally increases missed detections. It’s an inherent trade-off.

In traditional solutions, engineers can only pick a relatively comfortable working point along the ROC curve. To push both metrics down simultaneously, the only way is to shift the entire ROC curve lower—meaning the model’s discriminative ability must fundamentally improve.

NIO’s answer: the World Model’s long-horizon prediction ability—in short, decisions are based not on “is this object dangerous in this frame” but “what is the likelihood this object’s behavior will be dangerous in the next 3 seconds.” The more accurate the prediction, the more flexibility in choosing a working point.

Here, the World Model indeed has a structural advantage over pure perception-based end-to-end—it's inherently designed to “imagine the future.”

A Side Note: Why Automakers Are Rushing to World Models

By the second half of 2024, the World Model has basically become politically correct in the intelligent driving community—NIO NWM, Li Auto MindVLA, and the next-gen Xpeng XNGP are all converging in this direction.

The core reason: the end-to-end approach hit a ceiling. Pure imitation learning (IL) relies on massive driving video datasets for behavior cloning, but the corner cases on the road are endless. Worse, IL suffers from distribution shift—when the model veers slightly off during driving, it enters a state space never seen in training data, drifting further off course.

The World Model’s solution: “driving in your head.” Break the driving process into a “simulate → evaluate → act” cycle, letting the model run long-tail scenarios in a simulated world. Essentially, it uses generative models to create data to feed RL.

NIO’s advantage lies in having built NSim, a generative simulation model, early on, plus real-world scene data fed back via swarm intelligence, forming a relatively complete data flywheel. The ability for the SFT layer to finely sculpt behaviors is also predicated on having abundant, clean, high-quality human driving data.

A Small Assessment

In this update, “direct steering and pedal output” is the technical detail most worth watching, because it truly changes the control paradigm for intelligent driving. SFT’s addition to the training framework is an engineering-sound choice—not dazzling, but on point. Overhead sign recognition is more of a product-level differentiator.

Simultaneously upgrading 700,000 users, allowing four-year-old cars to benefit too—this is a relatively restrained and dignified approach among domestic manufacturers. That said, NIO’s real challenge is the stretch from “human-like assisted driving” to “genuinely letting go of the wheel”—whether the SFT and RL combo can bridge that gap will only be known next year.

As an aside, for developers wanting to do multi-model comparisons and connect intelligent driving model inference outputs with large language model planning layers, OpenAI Hub (openai-hub.com) lets you call GPT, Claude, Gemini, and DeepSeek with a single key—direct from China, OpenAI-format compatible, removing the hassle of integrating multiple different SDKs.

References

• NIO World Model New Version Begins Rollout: Over 700,000 Users Upgrade, 4-Year-Old Cars Included – IT Home: First-hand official rollout announcement on June 18
• NIO ES9 Officially Announces First Deployment of New World Model Version, Rolling Out to Banyan in June – IT Home: May preview detailing three-layer framework and end-to-end kinematic modeling background
• Practical and Reliable: “NIO World Model NWM” New Version Officially Rolled Out – Zhihu Column: Official NIO technical interpretation, including the “World Model + Closed-Loop RL” R&D paradigm explanation