[AI]+[Progress]: TaskMind General Task Planning System Enables Humanoid Robots to Complete 100 Unstructured Household Tasks

Article Body

Home environments are among the most challenging operational scenarios for robots. Unfixed object locations, vague task instructions, and dynamic environmental changes render mature technical solutions from industrial robotics almost entirely ineffective in household settings. The TaskMind system, jointly developed by Figure and OpenAI, is attempting to break through this bottleneck.

In May 2030, Figure held a TaskMind technology demonstration at its Sunnyvale, California headquarters. During the demonstration, a Figure 03 humanoid robot completed 100 household tasks in a completely unfamiliar apartment, including "collecting scattered clothes from the living room into the laundry basket," "finding salad ingredients from the refrigerator and washing and cutting them," and "making the bed and tidying up."

TaskMind's technical core is a "hierarchical task decomposition" architecture. When the robot receives a high-level instruction (such as "prepare a simple dinner"), the system first calls a large language model to decompose the instruction into executable subtask sequences, then uses a vision-language model for three-dimensional semantic understanding of the current environment, and finally generates specific joint angle trajectories through the motion planning module.

Mark Chen, VP of applied research at OpenAI, said: "TaskMind's key breakthrough isn't in any single module but in the overall system robustness. When the robot encounters unexpected situations during execution (such as knocking over a cup), it can autonomously assess the damage and adjust subsequent plans."

Out of 100 tasks, TaskMind achieved a success rate of 87%. Failure cases were mainly concentrated in fine motor tasks (such as threading a needle) and scenarios requiring understanding of implicit human intent (such as the specific standard for "tidying up the living room" varying from person to person).

Figure founder and CEO Brett Adcock revealed that TaskMind's training data came from real-machine collection in 2,000 home environments, totaling over 100,000 hours. Each home was equipped with 8 cameras and 12 force sensors to capture full-body movements and contact force data during human household tasks.

However, TaskMind still has a way to go before commercial deployment. A single Figure 03 unit costs approximately $250,000, with battery life of only 4 hours, and limited maneuverability in confined spaces (such as bathrooms). Figure plans to launch a B2B version for nursing homes and assisted living facilities in 2031, with no consumer version timeline announced yet.

Siddhartha Srinivasa, robotics professor at Carnegie Mellon University, commented: "TaskMind has moved humanoid robots from lab demonstrations into real-world validation. This is a breakthrough from 1 to 10, but getting from 10 to 100 still requires solving three major problems: cost, safety, and reliability."