Summary of Key Points
This conversation focuses on Jiejia Vision's approach in the field of embodied intelligence (robots that can physically interact with the world just like humans). The main points discussed include five key areas:
1. Model Approach: They choose a combination of a World-Action Model (WAM) and a language model, rather than settling for one or the other.
2. Current Industry Challenges: The biggest obstacle in the industry is not the model architecture itself, but the lack of high-quality data and a standardized evaluation system.
3. Building the Robot Body: Jiejia Vision insists on developing their own robot bodies to obtain real-world data and create a closed loop of model-body-scenario interaction.
4. Testing for Home Use: They start by testing home robots in standardized environments such as talent apartments before gradually introducing them to ordinary households.
5. Commercialization: The key to success is making users willing to pay for the robots, which means solving real household problems, not just demonstrating their capabilities in laboratories.
Detailed Explanation
1. WAM and VLA: Partners, Not Rivals
You can think of WAM as the “action executor” and VLA as the “task commander”:
- WAM’s strengths: It learns specific action details by watching videos (e.g., how to grip a pen or the force needed to cut tomatoes), making it suitable for precise tasks that require adaptation to different physical environments (e.g., picking up medicine from various pharmacy shelves).
- VLA’s strengths: It breaks down complex tasks into manageable steps using natural language (e.g., “cutting tomatoes, pouring oil, frying tomatoes, and adding eggs” for a recipe like “tomato and egg stir-fry”).
- How they work together: VLA divides large tasks into smaller steps, and WAM handles the precise execution of each step. For example, when preparing tomato and egg stir-fry, VLA instructs “cut the tomatoes first,” and WAM ensures the task is completed accurately, regardless of the placement of the ingredients.
2. Data and Evaluation: More Critical than Models
Many companies in the industry are debating model architectures, but Jiejia Vision believes that data and evaluation methods are the real challenges:
- Quality over quantity: Data acts as the robot’s “practice exercises” and must cover a wide range of scenarios (e.g., different kitchen layouts) with accurate action details (e.g., whether the robot touches nearby objects while picking up a cup). They estimate that home robots need about 10 million hours of high-quality data to function effectively, but they are far from that goal.
- Standardized evaluation: Similar to standardized exams, there needs to be a unified evaluation framework so that results from different teams can be compared fairly. Jiejia Vision has a dedicated team for automated scoring to minimize subjective biases.
3. Developing the Robot Body
While many AI companies focus on models, Jiejia Vision also develops the physical hardware (the robot’s body) for two reasons:
- Model validation: The effectiveness of the model can only be verified using their own hardware. Using other robots may lead to uncertainty or risks due to differences in platforms or supply chains.
- Real-world feedback: By deploying their robots in real environments (such as factories and talent apartments), they can gather valuable data on usage issues (e.g., dropping objects, users requesting the wrong item). This feedback helps improve both the model and the hardware, creating a continuous cycle of improvement.
4. Testing in Semi-Realistic Environments
Ordinary homes are too complex (with varying item placement and decoration styles), so Jiejia Vision starts with more standardized settings like talent apartments and hotels:
- For example, their Maker H01 robot is used in factories to test its precision and collect data.
- Once the model performs well in these environments, it can be adapted for ordinary households.
5. Commercialization
A home robot must be practical and meet two criteria to be successful:
- User acceptance: Users should be willing to buy and use the robot, perhaps requiring initial training but with gradually improving performance (e.g., automatically organizing clothes daily).
- Sustainable value: Users should see a clear benefit in paying for the robot’s services (e.g., monthly fees for cooking or cleaning assistance).
Jiejia Vision has delivered over 100 Maker H01 robots so far, with a goal of reaching 1,000 units this year. However, they are still evaluating the product’s reliability and customer satisfaction.
Conclusion
Jiejia Vision’s approach is pragmatic: they avoid debating theoretical model approaches and focus on practical solutions. They don’t aim for immediate market entry but instead test their robots in semi-realistic environments first. They prioritize data collection and evaluation, as these are essential for the successful implementation of home robots. For users, the key to reliable home robots lies in the quality of these foundational aspects.