In Search of Possibility

Jing Chen outdoors in a winter landscape

Jing Chen

Born in 1991 · Guangyuan, Sichuan

Physical AI System Engineer

  • Physical AI
  • Autonomous Driving
  • Systems Engineering

“I think this may be the essence of R&D:

to imagine beautiful things that do not yet exist,and bring them into the world.”

2024—2026, Chief Expert for ADS at Huawei Automotive BU, responsible for ADS 4.0 and its end-to-end AI solution, completing the transition from rules to end-to-end AI. 2016—2024, a founding member of DJI Automotive, where I built the planning and control algorithms department, the systems department and the intelligent R&D structure from the ground up. In 2020, we deployed China's first map-free driving system. 2014—2016, research on autonomous drone flight with safety guarantees at HKUST. 2010—2014, undergraduate study in computer science at Harbin Institute of Technology, including AI research at Microsoft Research Asia. In 2013, I began training language models on a cluster using the full Wikipedia corpus.

I recently started a venture focused on deploying physical AI systems at scale in real-world settings, opening up the possibilities of the physical world.

Experience & Ideas

2010—2014HIT & MSRA: How can AI Computing scale?

During my undergraduate years, we competed in ACM-ICPC programming contests, exploring how to measure computational efficiency and make better use of limited computing resources.

In 2013, while researching language models and named entity recognition (NER) at Microsoft Research Asia, we pursued a further question:

Can computing power become a scalable capacity to process information?

Yes. This is AI Computing.

Scaling AI Computing involves three core questions:

Compute: How can computational resources scale?

Clusters.

Organising and scheduling many nodes makes it possible to expand computing resources. We used the Cosmos cluster, with roughly 2,000–3,000 nodes, for large-scale data processing and model training.

Data: How can the supply of data scale?

Internet data.

The growing volume of natural-language material on the internet provides an expanding information source. We used the full Wikipedia corpus for our language-model and NER research.

Models: Can capability grow with scale?

Neural networks.

In our experiments at the time, methods such as SVMs and decision trees did not scale as well as we hoped. MLPs showed the potential to absorb more data and learn more unified representations, reducing the need for task-specific design. This pointed toward a path for model capability to grow with data and compute.

I have always wanted AI Computing to work in the physical world. I subsequently brought this research into robotics.

2014—2016HKUST RI: What makes the physical world difficult?

At HKUST, I studied real-time autonomous planning for drones: how could a drone keep flying safely with only a limited understanding of its surroundings? This led me to three fundamental questions about physical systems.

What makes physical systems harder than systems in the digital world?

The central difficulty is uncertainty, and how it propagates and accumulates over time.

  • Dynamic uncertainty: the world itself changes.
  • Observation uncertainty: a sensor's distance measurement can differ from the true distance.
  • Execution uncertainty: the same turning command can produce different actual turns.
  • Accumulation over time: small errors at each step affect subsequent position estimates and decisions.

Trying to walk in a straight line with your eyes closed illustrates the problem. Every step may introduce a small deviation. Without visual observations to correct it, the longer you walk, the harder it becomes to know where you are.

How can we handle uncertainty from multiple sources as it accumulates?

The core approach is redundancy to manage uncertainty.

  • Reduce uncertainty at its source: choose more stable sensors and motors with more consistent responses.
  • Fuse information: use hardware synchronisation to align observation times so different sources can work together.
  • Control in real time: use frequent observations and feedback to correct direction and speed promptly.

When riding a bicycle, your eyes observe the road, your body senses its tilt and your hands keep adjusting the handlebars. Changes in the road, judgement errors and small execution errors are always present. Multiple sources of information and continuous feedback let you keep correcting them and maintain balance.

If intelligence means making the right decisions, how do we make them?

At the heart of decision-making is anticipating the consequences of different choices.

After collecting lunch in a busy canteen, you carry a tray while looking for a seat, as people approach or suddenly stand up and pull out chairs. You anticipate where to pass, when to pause and whether the seat you spotted will still be free. As people move and seats fill, you keep adjusting your route to reach a suitable place.

  • Action space: how sharply a drone can turn and how quickly it can stop determine which actions remain feasible.
  • Predicting consequences: where will it go after moving forward, taking a detour or slowing down? What will it encounter, and how far will it be from its goal?
  • Real-time computation: effective models and data structures make it possible to explore multiple consequences further into the future in time to act.

The final choices are where to fly, how fast, when to turn and when to stop, in service of the task. My criterion is that uncertainty remains bounded, and the system's ability to act covers that range.

Research outcomes

  • Research and publications on autonomous drone flight with safety guarantees.
  • Improving Octree-Based Occupancy Maps using Environment Sparsity with Application to Aerial Robot Navigation later became a finalist for the ICRA 2017 Best Service Robotics Paper Award. Official list ↗
  • Our drone obstacle-avoidance experimental setup became a de facto academic standard.
2016—2024DJI Automotive: How do we reason and decide amid complexity and uncertainty?

Why is autonomous driving difficult?

Uncertainty and complexity: autonomous driving requires handling uncertainty within a complex system.

Uncertainty comes from the physical world: observations can be incomplete, and other participants' behaviour cannot be fully predicted.

Complexity comes from the many elements of a driving environment and their interactions. They increase the dimensions of the problem and the number of scenarios that need to be understood and validated, leading to Bellman's “curse of dimensionality”:

  • Static space: road boundaries, road geometry and drivable areas.
  • Semantic rules: lane attributes, travel directions and right of way.
  • Dynamic objects: vehicles, pedestrians, their movement and their interactions.

Why is map-free driving even harder?

An additional dimension of uncertainty: the road structure itself must be understood and assessed in real time.

Systems that depend on HD maps typically treat the road structure supplied by the map as known, effectively removing that dimension of uncertainty from the design. A map-free system must infer the road structure while driving and deal with the incompleteness and uncertainty of that understanding.

On your first drive in Shanghai, elevated roads branch repeatedly, signs are dense and turning lanes may not be where you expect them. Approaching a junction, the vehicle ahead may also hide the lane arrows. Even with a navigation route, you must interpret the road connections and lane destinations from incomplete information and choose in time.

How did we address complexity?

System architecture and engineering: this became the foundation for how I built the R&D organisation and systems engineering structure, with three elements:

  • High cohesion, low coupling: clear module boundaries and interfaces allow specialists to develop their domains while reducing the complexity they need to handle outside them. Dedicated systems experts take responsibility for cross-domain integration.
  • Engineering accumulation: build reusable units of capability and standardise interfaces within a clear division of responsibilities. Deployment and feedback continually add benchmarks, testing capabilities, understanding and engineering know-how that can inform later work.
  • Reliability through redundancy: prepare alternative information sources and responses for missing information or limited capabilities, allowing the system to continue working when a local source or capability is unavailable.

Replacing a household appliance within the rated supply and load limits only requires connecting it through a standard interface. Its internal power and control circuits handle its own operating requirements, while development and testing can be organised around the agreed interface. This makes it possible to replace a component without redesigning the building’s power system, and to combine and reuse encapsulated capabilities.

How did we solve the map-free problem?

Multi-source information fusion: to address the additional uncertainty about road structure, we combined three sources:

  • Lane markings: their positions and directions provide direct evidence of road structure.
  • Other vehicles' behaviour: when vehicles obscure the markings, surrounding trajectories help infer the road's direction.
  • Road boundaries: when clear markings and nearby vehicles are both absent, boundaries help identify drivable space.

Roadworks cover the original lane markings, while vehicles ahead obscure part of the road. The system can still combine surrounding vehicle trajectories, road edges and temporary barriers to infer the current drivable space. These sources complement one another when individual observations are missing.

How can a system make good decisions in an uncertain environment?

Defensive driving: assess the feasible action space while accounting for uncertain observations and the uncertain behaviour of others when predicting consequences.

On narrow streets in Puxi, parked vehicles and building corners can block the view. A pedestrian may be about to cross, or a bicycle may emerge from an unseen area. The system needs to include these unobserved possibilities in its predictions, evaluate the consequences of different actions and retain room to respond.

  • Action space: identify the feasible actions and include them in the evaluation.
  • Uncertain observations: a blind spot may contain a vehicle or pedestrian that has not yet been observed; consider both possibilities.
  • Uncertain behaviour: another vehicle may or may not cut in; a pedestrian may step out or remain still.
  • Consequences across possible futures: combine these uncertainties to anticipate multiple ways the world could evolve, evaluate each feasible action under those possibilities, and then choose an action.

What did this work achieve?

Technical deployment and commercial outcomes:

  • Deployment: I was responsible for China's first deployed map-free driving system, delivered in 2020, three years ahead of the wider industry.
  • Commercial outcome: we secured China's first major autonomous-driving contract: a million-unit order for Volkswagen's RHP programme, contributing an important business result for the company.
2024—2026Huawei Automotive BU: How can physical AI systems combine scaling laws with safety?

A mathematical view of uncertainty, implemented through generative models.

  • Hyperspace: representing the world's possibilities. I view all possible world states and behaviours as a high-dimensional space that the model first needs the capacity to represent.
  • Embedding space: learning the data distribution within that space, capturing the states and behaviours supported by the available information.
  • Value alignment: choosing behaviours within the data-supported space according to product values and safety objectives.

How can physical AI systems exhibit scaling laws?

Generative pretraining: first enable the model to represent possibilities, then learn the data distribution.

  • How can a model represent the world's possibilities?

    Use generative models to open up the representation space.

    A central limitation of earlier approaches was insufficient representational capacity: they could not effectively represent many behaviours present in real data. The first step was to overcome that limitation so the model could express richer states and behaviours within the hyperspace.

  • How can it learn the data distribution within this representation space?

    Self-supervised learning with a cross-entropy objective.

    Earlier approaches supervised multiple complex objectives within the same learning process, mixing the task of learning the data distribution with other goals. My approach was to move those other goals into post-training, allowing pretraining to focus on the data distribution and capture the data-supported embedding space.

Along this path, in 2025 I was responsible for a single-stage, end-to-end solution, using data-driven learning to replace hand-written rules in the main pipeline and directly output trajectories for vehicle control.

How do we achieve safety?

Value alignment: teach the model to choose behaviours according to product values and safety requirements.

When merging into slow traffic, real driving data includes both waiting for a suitable gap and forcing into a small one that makes the following vehicle brake sharply. A model encounters both behaviours in the distribution, but not every observed behaviour is appropriate for the product. Post-training must guide its choices according to safety, comfort and mobility objectives.

We introduced product values and safety objectives during post-training, using on-policy reinforcement learning for alignment, together with a risk field to build defensive-driving capability. Pretraining learns the data distribution; post-training learns how to choose behaviours on that foundation.

How do we ensure reliability?

Through system redundancy.

Different physical mechanisms can provide vehicle deceleration. If a friction-braking channel becomes unavailable while the electric drive and energy-recovery path remain operational, regenerative braking may retain some deceleration capability under suitable operating conditions. Redundancy means preserving available ways to respond when one part fails.

AI systems can fail. In a complex system such as autonomous driving, I positioned the traditional rule-based system as a parallel fallback for the AI main system, intervening when the main system fails. This avoids continuous interference during normal operation that could limit the model's ability to scale. It is also the end-to-end main system plus rule-based fallback architecture I pursued toward L3 and L4.

What did this work achieve?

The end-to-end solution delivered more human-like driving and an order-of-magnitude improvement in safety. ADS 4.0 reached million-scale vehicle installations, with a subscription rate above 90% and an average share of user driving mileage on intelligent driving also above 90%.

At the China EV100 Forum, the company also announced that it would increase investment in AI computing power fivefold, continuing its commitment to capability growth driven by data and compute.

2026—Next: How can physical AI enter more real-world settings? How can it explore the unknown and open up possibilities?

Where does the long-term commercial value of physical AI systems lie?

Deployment: building engineering experience in real-world settings. This accumulated experience cannot be replaced by codified knowledge alone, and will become a lasting source of knowledge and information for AI.

Why won't AI replace this accumulated engineering experience?

The essence of intelligence is the generalisation of capability, rather than overfitting to existing information.

The same robotic arm, controlled by models with different movement patterns, can develop different heating, wear and accuracy characteristics while performing the same picking task. As component behaviour drifts, the model’s observations, the actual outcomes of its actions and its subsequent choices can also change. Engineering records must therefore identify the complete hardware and software configuration, model version and operating conditions to establish which combinations work together over time.

For system-level engineering information accumulated over time, the optimal path intelligence has found through its evolution is information retrieval plus in-context learning. Validated configurations, operating conditions and long-term performance records are stored in an external knowledge base, where AI can retrieve, understand and use them for the current task without repeating the memorisation or validation.

Why is engineering experience so important?

System-level non-functional properties require validation across operating conditions and over time.

A few successful picking demonstrations with a new robotic arm cannot establish the complete system’s long-term durability, thermal drift or accuracy retention. These non-functional properties require evidence from accumulated operation, lifetime testing and failure analysis, including validation of how the model responds when components drift. Storing that evidence in a knowledge base with its operating conditions, configuration and versions lets later AI systems reuse the validated experience.

Volkswagen and Toyota’s accumulated experience in manufacturing, supplier coordination and fault handling supports reliable products and production at scale. Toyota’s quality history ↗ Policy support, sustained investment and large-scale use have also built manufacturing experience, industrial coordination and supply-chain advantages in China’s EV industry. IEA automotive industry report ↗ These capabilities come from long-term practice, and talented people and advanced ideas still need to be tested through production, deployment and feedback.

How can AI systems autonomously explore the unknown and open up possibilities for people?

Meta-capabilities: understanding which actions a system can take, and how those actions affect its own knowledge, capabilities and uncertainty, so it can explore and grow autonomously.

  • How can AI distinguish the known from the unknown and respond differently?

    Metacognition: recognise the boundaries of its knowledge and capabilities, distinguish supported judgements from questions that lack information, and align actions with what it knows and can do.

  • How can AI actively change its own information and capabilities?

    Meta-actions: understand how actions change its information, capabilities and uncertainty, and choose actions that acquire information, accumulate experience and expand its capabilities.

    A user asks the robot to bring their glasses, but the robot does not know where they are. It uses the available information to decide whether to check the desk, bedside or storage area first, updating its location estimate after each observation and asking the person when needed. By actively gathering information, it narrows the unknowns until it can complete the task.

  • How can AI improve its ability to learn in the face of the unknown?

    Meta-learning: continually test and improve its learning methods through action and feedback, becoming better at gathering information, updating its understanding and developing capabilities when faced with new unknowns.

    A user asks for a dish the robot has never cooked, so it needs to find a method, understand the steps and test and adjust through actual preparation and feedback. After learning it, the robot continues adapting to the person’s taste, seasonal ingredients and changing food interests. It can also improve its own learning methods through that feedback, continually expanding its repertoire and ability to meet new needs.

  • Where do information and value for autonomous exploration come from?

    The real world: it is a source of both information and value. Systems must be deployed in the real world to obtain data and feedback through interaction, and find direction for exploration and evolution in real needs.

I am committed to enabling AI to explore the unknown in real-world settings and open up possibilities for people.

Résumé

2024—2026

Huawei Automotive BU

Chief Expert, ADS

Responsibilities

  • Responsible for the end-to-end technical solution, coordinating multiple development teams and organising technical problem-solving
  • Responsible for the ADS 4.0 solution and ADS 5.0 technology planning
  • Managed the systems engineering (SE) expert team and developed its AI capabilities
  • Coordinated technical planning across product development (N), next-generation research (N+1) and frontier research (N+2)

Key work

  • In 2025, responsible for the single-stage, end-to-end AI solution, replacing rules with data-driven learning and directly generating trajectories for vehicle control
  • Developed defensive driving through risk-field modelling and on-policy RL
  • Designed high-safety architectures toward L3 and L4, combining an end-to-end main system with a rule-based fallback

Outcomes

  • Global top-two performance for the single-stage end-to-end solution in 2025
  • Million-scale vehicle deployment of ADS 4.0
  • User subscription rate above 90%
  • Intelligent driving accounted for over 90% of user driving mileage

2016—2024

DJI Automotive

Founding member · Head of Planning & Control and Systems Departments

Responsibilities

  • Built the planning and control algorithms department from the ground up
  • Built the systems department from the ground up
  • Established a matrix R&D structure

Key work

  • Responsible for China's first deployed map-free driving system (2020), three years ahead of the wider industry
  • Built defensive-driving capabilities to handle multiple forms of uncertainty
  • Built a platform architecture with reusable capability units and standardised interfaces, supporting continued accumulation of capabilities around shared interfaces

Outcomes

  • Deployed China's first map-free driving system in 2020, three years ahead of the wider industry
  • Secured a million-unit order for Volkswagen's RHP programme
  • Secured vehicle programme orders from 50 other OEMs

2014—2016

Hong Kong University of Science and Technology

Robotics Institute · MPhil, full scholarship

Research focus

  • Autonomous aerial systems
  • Safety and real-time performance during autonomous operation

Key work

  • Pioneered corridor-based planning with safety guarantees
  • Pioneered real-time online autonomous mapping based on octrees
    Technical details: Octree query and update complexity
    Volume occupancy query · Upper bound
    CPUO((vr/Vε)2/3)
    GPU (fully parallel)O(log(vr/Vε))
    Single-ray map update · Worst case
    CPUO((Lr/Vε1/3) · log(vb/Vε))
    GPU (fully parallel)O(log(vb/Vε))

    vr is the query volume; Vε is the finest voxel volume; Lr is the ray length; vb is the volume of the ray’s axis-aligned bounding box

  • Built the first autonomous flight system to run in real time on onboard computing hardware with safety guarantees

Outcomes