Generalist’s GEN-1 Robot Model Achieves 99% Task Success and Emergent Physical Reasoning

Generalist, a previously low‑profile Silicon Valley robot‑AI startup, has unveiled its next‑generation foundation model GEN-1, bringing large‑model techniques into the robot world.

Compared with the prior Gen‑0’s average 64% success rate, GEN-1 reaches 99% success across multiple tasks such as T‑shirt folding and vacuum maintenance, and can run hundreds of cycles without human intervention. For box‑folding, execution time dropped from 34 seconds to 12 seconds, roughly a three‑fold speedup.

The most notable advance is the emergence of “physical commonsense.” In a lengthy automotive‑part assembly, when a washer becomes misaligned, the robot either re‑places it, adjusts its pose using the surrounding structure, or employs both hands to secure it. For soft, deformable objects that assume unexpected shapes, the model autonomously restores them to an operable state—behaviors that were not present in the original training data.

GEN-1 was trained on more than 500,000 hours of human‑captured data using a “data‑glove” that records visual and motion information as people perform tasks in real environments like homes and warehouses. CEO Pete Florence links this scaling of model size and data to the sudden capability releases seen in early ChatGPT development.

This approach marks a strategic shift: robots are no longer treated as highly customized mechanical systems but as continuously trainable model platforms, with the goal of scaling models, data, and compute to approach general‑purpose ability.

The industry consensus is that robotics now lacks data more than models. Unlike the internet, the physical world does not provide a readily scrapeable data source for edge‑case handling. Generalist’s solution is to turn humans into data generators via the data‑glove, creating a massive, real‑world dataset.

Physical Intelligence, another prominent player, relies more on teleoperation and simulated environments for data collection, highlighting contrasting data‑generation strategies.

There is debate: experts such as Brad Porter warn that without mature underlying architecture, merely amassing data can be costly and ineffective, as history shows scaling must accompany architectural breakthroughs. Nonetheless, Generalist secured $140 million in 2025 funding at a $440 million valuation, backed by Spark Capital, NVIDIA’s NVentures, Bezos Expeditions, and Boldstart Ventures. Physical Intelligence is approaching a $10 billion valuation, and Jensen Huang has declared that robotics is entering its “ChatGPT moment.”

Viewed over time, this wave represents a paradigm shift from scripted task execution to data‑driven learning, and Generalist aims to prove that when model size, data volume, and compute cross a critical threshold, robots will experience a decisive capability leap.