Generalist AI Releases GEN-1.5: One-Shot Imitation Learning Model for Humanoid Manipulation

One-Shot Video Demonstration for Manipulation Policies

On August 30, 2026, robotics lab [Generalist AI] released technical results for [GEN-1.5 One-Shot VLA]. The system generates manipulation trajectories after processing a single video demonstration of 3 to 12 seconds.

Benchmark Metrics

Across 10 standardized manipulation benchmarks:

Zero-Shot Success Rate: 59.0% from a single video demonstration without task-specific code.

Fine-Tuned Success Rate: 83.0% with 5 minutes of supplementary demonstrations.

Inference Latency: 22 ms on edge compute hardware (NVIDIA Jetson AGX Orin).

Evaluation

In test trials, the system demonstrated functional tool substitution under missing object conditions (e.g. selecting an alternate object with similar surface contact characteristics to complete a clearing task).

Tournament Context

In autonomous leagues such as WHRG Autonomous Manipulation, Vision-Language-Action (VLA) models reduce setup time for unseen obstacle and tool configurations.