Mentee humanoids are engineered to deliver robust out-of-the-box capabilities. These include advanced scene understanding, natural instruction following, end-to-end autonomous task execution without teleoperation, and reliable locomotion, navigation, and safe manipulation of rigid objects.

In this post, we introduce two capabilities: mentoring and instruction following.

A. Mentoring

Observe → Train → Perform (Real2Sim2Real)

One of Mentee's most distinctive capabilities is few-shot generalization. This allows a MenteeBot to acquire and execute new skills even after a single human demonstration, enabling rapid deployment across a wide range of real-world tasks. This approach eliminates reliance on continuous teleoperation or large-scale real-world data collection.

The mentoring process follows three clear stages, which together form a Real2Sim2Real (Real-to-Sim and back to Real) sequence.

Observe

A MenteeBot observes a human mentor performing a task from the robot's own viewpoint as it watches the human operate in the environment. This external viewpoint captures the full-body motion of the mentor, the interaction with objects, and the overall task flow exactly as the robot perceives it in the real world. The demonstration is recorded as a video, without teleoperation, markers, or instrumentation.

Train

The recorded video is then ingested into Mentee's simulation environment using foundation models that interpret the visual demonstration and reconstruct the task inside the simulator, including object relationships, motion trajectories, and task structure.

Once instantiated in simulation, training is performed at scale, using Self-play and RL (reinforcement learning). The simulator enables thousands of task variations with small but meaningful differences, covering edge cases that would be slow, costly, or unsafe to collect in the real world. This process enables rapid learning with low marginal cost and strong generalization. At the moment, it runs offline, but we aim to make it instantaneous.

Perform

Using Mentee's proprietary Sim2Real technology, the trained behavior is transferred back to the physical robot. The MenteeBot then executes the task autonomously in the real world, preserving the intent of the original human demonstration while remaining robust to real-world variability.

The video below demonstrates a MenteeBot learning, through mentoring, how to replace a battery in another MenteeBot:

B. Instruction Following

One way to have a robot perform a task is through demonstration. Another is by giving an instruction.

Here, we demonstrate instruction-following capabilities. When asked to bring a can, the robot understands the request, identifies the correct can, navigates to the counter, searches for a matching item, picks it up, and brings it back to the person. This sequence combines language understanding, scene perception, navigation, and manipulation into a single coherent behavior.

This approach significantly lowers the barrier to human-robot collaboration, enabling non-experts to deploy humanoid robots effectively in real-world environments.

The following video demonstrates instruction following in action: