Tony Zhao Releases ACT-2 Preview Robot Model, Laundry Folding Success Rate 99.1%
Tony Zhao introduced ACT-2 Preview, the first robot model that unifies broad generalization and high reliability, capable of teaching Memo new behaviors and achieving generalization with a single fine-tuning example.
In 785 trials across 31 unseen environments, ACT-2 achieved a laundry folding success rate of 99.1%, with folding quality at human levels, an average score of 4.72/5, and 98.3% receiving four or five stars.
The model exhibits significant emergent capabilities: robustly picking up clothes from the ground, handling baby clothes to 8XL shirts, and resisting interference and lighting changes.
Tony Zhao defines the concept of "Solve": clarifying the scope and adapting cost boundaries, avoiding progress assessment based solely on demonstrations. ACT-2 achieves "Solve" through large-scale pre-training combined with a small amount of internal data.
Single-example fine-tuning allows for generalization of new behaviors, and significantly expanding the pre-training scale reduces the performance gap between in-domain and out-of-domain, making internal performance a reliable predictor of field performance.
The Memo robot fleet operates in parallel, rapidly enhancing reliability, quality, and speed.
Source: Public Information
ABAB AI Insight
Tony Zhao's team focuses on generalization and reliability in robotics, with ACT-2 achieving high success rates in complex tasks like laundry folding through large-scale pre-training and minimal fine-tuning. Single-example learning and fleet parallel iteration mark a shift from narrow demonstrations to broad "Solve" capabilities.
On the capital front, such high-reliability robot models will accelerate deployment in household and industrial settings, with funding concentrating on projects that demonstrate quantifiable generalization metrics and real-world production validation, further strengthening the competitive barrier of the Memo fleet data flywheel.
Similar robot startups are releasing milestone achievements, currently transitioning embodied intelligence rapidly from laboratory settings to home environments.
Essentially, this represents a technological replacement: ACT-2 replaces traditional task-specific programming with a unified generalization model, significantly reducing robot adaptation costs, shifting pricing power from hardware-centric to software-defined systems based on pre-training and fine-tuning, promoting the commercialization of household robots.
ABAB News · Cognitive Laws
Single-example generalization surpasses thousands of programming examples; scalability determines the ceiling for robots.
"Solve" defines scope and cost, with quantifiable metrics ending the era of vague demonstrations.
Fleet parallel iteration and data flywheels allow internal optimizations to directly translate into field reliability.