Open-source release includes 11,430 real-world trajectories spanning expert demonstrations, autonomous policy rollouts and human-in-the-loop corrections

【SHANGHAI, CHINA, September 1, 2026】 AGIBOT today announced the open-source release of AGIBOT WORLD 2026 Theme 3: Reinforcement Learning, expanding its embodied AI dataset initiative with real-world robot experience data collected across expert demonstrations, autonomous deployment and human-in-the-loop corrections.

The first release includes 11,430 real-world trajectories across 14 industrial and household tasks, including 1,024 successful policy rollouts and 1,369 failed rollouts. It also provides fine-grained annotations covering task progress, completion, errors, disturbances and human interventions, giving researchers structured execution feedback for reinforcement learning and related embodied AI research.

AGIBOT WORLD 2026 is an open-source dataset initiative built around real-world scenarios and multiple research themes, each with dedicated data-collection methodologies and annotation frameworks. Following earlier releases focused on imitation learning and diverse interactions, Theme 3 extends the dataset from learning primarily from demonstrations toward learning from real-world execution experience.

Large-scale expert-demonstration datasets have helped advance robot imitation learning by showing models how tasks can be performed. Real-world deployment, however, produces a much broader range of outcomes. Robots may succeed or fail, deviate from intended behavior, recover from errors or require human intervention.

These interactions reveal information that successful demonstrations alone cannot fully capture, including a model’s capability boundaries, failure states, effective behaviors and recovery processes. Theme 3 preserves these experiences so researchers can study not only how a task should be completed, but also what happens when robots attempt the task themselves.

From Demonstration to Deployment and Correction

AGIBOT WORLD 2026 Theme 3 brings together three complementary types of real-world trajectories: expert demonstrations, policy rollouts and human-in-the-loop corrections.

Expert demonstration trajectories record reference executions performed by human operators in real-world environments.

Policy rollouts capture autonomous model execution during deployment, preserving both successful and failed attempts. Successful rollouts show capabilities the policy has acquired, while failed rollouts help reveal capability boundaries and failure points at different stages of task execution.

Human-in-the-loop correction trajectories capture the model’s behavior before intervention, the point at which a human operator takes control and the corrective actions that follow.

Together, the three trajectory types form an end-to-end data pipeline spanning demonstration, autonomous execution and correction. Expert demonstrations provide reference behavior, policy rollouts show how models perform independently, and human-in-the-loop trajectories capture how execution errors can be identified and corrected.

Expert Demonstration

Policy Rollout

Human-in-the-Loop Correction

The first open-source batch focuses on industrial and household environments and includes 14 interactive task types, such as Ethernet-cable insertion and unlocking doors with keys. The tasks cover precision manipulation, long-horizon execution and contact-rich interactions.

In total, the first batch contains 11,430 trajectories, including 1,024 successful policy rollouts and 1,369 failed policy rollouts.

In motion-control Turning Real-World Execution Into Learning Signals

Reinforcement learning requires more than a record of what a robot did. It also depends on feedback about whether an action was effective, how far a task progressed and where errors, disturbances or interventions occurred during execution.

Theme 3 therefore provides fine-grained annotations covering task progress, completion status, errors, disturbances and human interventions. These annotations allow researchers to analyze execution at the process level rather than relying only on trajectory-level success or failure.

Fine-grained Annotation

The first batch contains 98,159 annotated subtask intervals, each with a completion-status label, together with 26,493 disturbance segments, 5,795 error-state segments and 10,684 human-intervention segments.

Researchers can use these annotations to develop more granular learning signals and support research into reward models, value models, task-progress prediction, success detection, error recognition, risk warning, intervention prompting and error recovery.

Data Applications

By combining shared real-world trajectories with structured execution feedback, AGIBOT WORLD 2026 Theme 3 aims to help reduce the cost of collecting real-robot execution data, enable research teams to reproduce experiments and compare methods using common trajectories, and turn experience collected across individual scenarios into a reusable resource for the broader research community.

From Demonstrations to Learning From Experience

Robots operating in dynamic physical environments need more than the ability to reproduce standardized actions. They must also be able to interpret feedback, recognize errors, recover from unexpected states and improve their behavior over time.

Imitation learning enables models to learn from expert demonstrations. Reinforcement learning provides a complementary approach by using execution feedback to evaluate which behaviors advance a task, which states indicate risk and how those observations can be translated into learning signals.

AGIBOT WORLD 2026 Theme 3 is designed to support this transition from learning from demonstrations to learning from experience. By preserving successful and failed executions, disturbances, human interventions and correction processes, the dataset provides researchers with real-world interaction data that exposes both what robots can already do and where their capabilities break down.

The release is part of AGIBOT WORLD 2026. AGIBOT plans to continue expanding the initiative with additional datasets, benchmarks and research resources for embodied AI.

AGIBOT WORLD 2026 Theme 3: Reinforcement Learning
Project page: agibot-world.com
Open-source dataset: huggingface.co/datasets/agibot-world/AgiBotWorld2026

For more information, please visit AGIBOT.com and follow AGIBOT on:
X: https://x.com/AGIBOTofficial
LinkedIn: https://www.linkedin.com/company/agibot/
YouTube: https://www.youtube.com/@AGIBOTofficial
Instagram: https://www.instagram.com/agibotofficial/
TikTok: https://www.tiktok.com/@agibotofficial
Facebook: https://www.facebook.com/AGIBOTofficial/

About AGIBOT
AGIBOT is a pioneer in the global general-purpose AI robotics industry, with its core technology focused on embodied AI. AGIBOT's "Three Intelligences in One" architecture integrates Locomotion Intelligence, Interaction Intelligence, and Manipulation Intelligence into a unified embodied system. Its portfolio spans humanoid robots, quadrupeds, dexterous systems, and commercial cleaning solutions. In June 2026, AGIBOT announced that its 15,000th robot had rolled off the production line.