AGIBOT's GO-3 Raises Retail Robot Task Success from 28.6% to 85.7% in Five Days Through Autonomous Learning

The Live Deployment

AGIBOT sent a G2 humanoid robot into an operating convenience store and tracked its task completion rate without any outside intervention. On day 1, the robot finished the target task 28.6% of the time. By day 5, the success rate reached 85.7%.

The improvement came from AGIBOT's Genie Master, an agent framework that connects decision-making, task execution, and feedback in a closed loop. When a task fails, the system identifies the cause, adjusts the approach, and commits the corrected strategy to memory for the next attempt. No engineer touches the model between tries.

A parallel component called Genie Evolver runs distributed training across roughly 100 physical robots simultaneously, blending real-world operational data with physics simulations and world models. Each robot's experience feeds the shared pool, accelerating improvement across the fleet.

The Architecture Behind It

GO-3 Preview runs on WRAM, short for World Reasoning Action Model. WRAM integrates 3 functions into a single framework: interpreting task instructions, anticipating the physical consequences of a planned action, and generating the motor commands to carry it out.

The practical result is that robots running GO-3 generalize beyond the exact sequences they've seen before. In a separate evaluation on hook-hanging tasks, GO-3 improved 17.7 percentage points over the conventional supervised fine-tuning baseline.

AGIBOT's training pipeline spans 3 stages: pre-training on first-person human video and robot trajectory data, post-training for specific deployment environments, and continuous improvement once the robot's working live. The company calls this Physical Intelligence Scaling.

Three Leaderboards at Once

GO-3 Preview took top positions on 3 embodied AI benchmarks in the same week as the convenience store results were disclosed.

On RoboDojo, a simulation-and-real benchmark spanning 42 simulated and 18 real-world manipulation tasks, GO-3 ranked first in task success rate and process score. On RoboTwin, a dual-arm manipulation leaderboard, it posted the highest overall score. On RoboColiseum, which scores instruction following, spatial understanding, robustness, and general-purpose manipulation, GO-3 ranked first in overall score. According to AGIBOT's official release, the company's companion omni-modal model, WITA-Omni, ranks first on the Daily-Omni embodied interaction benchmark.

The Company and What's Next

AGIBOT announced GO-3 Preview at AGIBOT Innovation Day 2026 in Shanghai. The company produced its 20,000th robot in September 2026, and its ecosystem now spans 255 universities, 511 enterprise partners, and more than 80,000 developers.

At the same event, AGIBOT introduced Worken, a productivity metric the company defines as Active Robots × Autonomous Working Hours × Output Efficiency Ratio. Peng Zhihui, co-founder and CTO, framed the shift plainly: "How much productive work are they actually delivering?"

AGIBOT plans to open-source the full GO-3 model before the end of 2026.

Sources

• Humanoid robot achieves 85% task success rate with AGIBOT's GO-3 model, Interesting Engineering
• AGIBOT Introduces Worken to Measure Productive Robot Work at AGIBOT Innovation Day 2026, PR Newswire
• AGIBOT introduces Worken to measure robot productivity at Innovation Day 2026, The Robot Report
• AGIBOT introduces Worken to measure productive robot work at Innovation Day 2026, Robotics and Automation News
• AGIBOT Innovation Day 2026 Official Announcement, AGIBOT
• AGIBOT Introduces Worken, PR Newswire APAC

Companies mentioned

• AGIBOT