EngramTeleOp
Vision-only algorithms translate bare-hand movements into robot control. Ultra-low-latency wide-area teleoperation records task video, commands and robot state in sync.
Explore EngramTeleOpTranscEngram 忆生科技
Autonomous intelligence for machines
Starting from the first principles of intelligence, we model how people learn and remember. Our aim is to connect perception, memory, prediction and interaction in a continuous loop, building a form of “silicon-based life” that can correct itself and evolve through experience in the open physical world.
Turning each interaction into experience a machine can use again.
In our research, a machine’s memory takes the form of two models. One describes how the physical world changes. The other describes how the machine’s own body moves. Both can be updated as the machine acts and observes the result.
Perception
Observe the surroundings and the machine’s own state to establish what is happening now.
Memory
Relate these observations to past actions and their outcomes. Use that experience to update the world and body models.
Prediction
Use the models to anticipate how a possible action would change the machine’s state and its surroundings.
Interaction
Act and observe the outcome. Differences between the prediction and the observation feed back into the next update.
Foundations for AI 2.0
Data foundations and world models
Connecting perception, memory, prediction and interaction takes real-world experience and environments where actions can be tested. Teleoperation and first-person capture build the data foundation; manipulable world models support scene reconstruction and physical interaction. Together, they provide the foundations for learning, validation and continued improvement.
EngramEgo
First-person input and full-body reconstruction
First-person data capture and generative 4D hand–object reconstruction. Recover how hands and objects move and interact in 3D over time, turning real demonstrations into training data.
Explore EngramEgoWorld Model
World perception models reconstruct scenes and object structure. World action research studies reusable skills, policy generation and generalization to new tasks.