"Zero-Added" Model: Zero Manual Data Collection, Zero-Shot Deployment, Zero Remote Intervention.
From Human Video to Proactive Service—No More Hand-Holding.
This round of training involves no teleoperation trajectories as training data throughout the entire process. All tasks were accomplished autonomously by the model itself—handling the unexpected is the true test! As shown in the videos, the model demonstrates online adaptation and replanning capabilities in unanticipated human-robot and multi-robot interaction scenarios. The resounding success of this banquet hall service is fundamentally attributed to the model's autonomous learning, exploration, and self-evolution capabilities.
Gone are the days of the ten-million-hour data gap. Instead of requiring task-specific customized training, the model extracts intent and tool-use commonsense knowledge from just 300 human videos. Consequently, when deployed in a novel environment, it autonomously analyzes tool properties and affordances, and explores to formulate new tool-use strategies. On this basis, it leverages human behavioral cues and its intent understanding capabilities to deliver proactive service—this is what we call "proactive intuition."
With no pre-set scripts, the model operates entirely in real time, navigating dynamic and uncertain interactive environments while responding to ad-hoc human requests. It performs in a single, continuous take—no cuts, no retakes, and no teleoperation interventions.
Empower the model to "perceive, reason, experiment, take initiative, collaborate, and evolve." Make Veritable People!
What enables the model to perform in a single, continuous take?
AIntent, Affordance, and Self-Evolution
The model constructs a cognitive system that is energy-driven, stabilized by attractors toward goals, proactive in information acquisition, and capable of behavioral correction through interaction, while adaptively allocating cognitive resources according to task difficulty. It goes beyond merely understanding the environment, predicting the future, and outputting actions.
How do attractors facilitate intent understanding?
The goal energy is defined as:
where denotes the current latent state and represents the goal state. A lower energy value indicates that the current state aligns more closely with the desired goal.
Attractors enable the system to maintain goal stability, suppress irrelevant states, and facilitate the activation of relevant memories or skills. Meanwhile, energy serves as a metric to evaluate the plausibility of candidate states and actions, enabling multi-step iterative reasoning, thereby achieving proactive intent through the following cycle: generating intent hypothesesassessing uncertaintyacquiring valuable informationupdating intentsdetermining actions.
Affordance—An Intermediate Representation Bridging World Understanding and Action Decision-Making
This is a task-conditioned affordance inference integrated with an energy-driven approach:
where:
- the discrepancy between the action outcome and the task goal
- the degree of violation of physical constraints
- risks such as collision, dropping, or damage
- the uncertainty of action outcomes
In our model, affordance is no longer merely a recognition label, but rather a basis for decision-making that actively participates in intent inference, planning, skill selection, and action refinement.
Adaptive Reasoning—Brief Thinking for Simple Tasks, Extended Thinking for Complex Ones
Based on internal state dynamics, goal discrepancies, prediction errors, and uncertainty, the model determines whether to continue or terminate reasoning. Within this process, attractors govern the system's trajectory: maintaining goal stability and activating relevant memories and skills to enable proactive perception.
Why does the model exhibit subconscious-like emergent behaviors?
AEnergy-Driven "Proactive Intuition"
- Why SubconsciousnessEnergy-Driven Mechanisms
- Through energy evaluation, attractor dynamics, and internal state regulation, the model develops implicit judgments, automatic behavioral tendencies, and stable goal-directedness, eliminating the need to rely on explicit reasoning at every step.
- Why EmergenceInformation Gain
- The model proactively acquires information, explores the environment, and learns from the consequences of interactions, enabling complex behaviors to emerge from foundational capabilities and continuous experience.
- Why Proactive IntuitionAngular Gyrus, Social Cognition, and Proactive Intent Understanding
- By integrating contextual understanding and Theory of Mind, the model infers humans' goals, beliefs, and emotions, comprehends unexpressed needs, and determines when to take proactive action, inquire, or wait.
How is the model capable of self-learning?
ASelf-Evolution and Multimodal Data Engine
Why "Fill in the Blanks"—Predictive Completion?
Ah! Someone finally noticed that robots don't necessarily need wrist-mounted cameras. When information is missing or occluded, the model leverages constrained inference—grounded in world models, memory, attractors, and energy evaluation—to perform predictive completion. This capability comprises three components: perceptual completion, contextual completion, and intent completion. Attractors provide a stable state-organization mechanism, while energy functions serve as a hypothesis-evaluation mechanism. Integrated with information gain, these elements form a verifiable, energy-driven closed-loop completion system. The loop operates as follows: Information GapHypothesis GenerationEnergy EvaluationUncertainty AssessmentContinued Reasoning or Active ObservationHypothesis Update. Definition:
where denotes the expected information gain of an action, represents the temporal or energetic cost, and indicates the associated risk.
Why No Real-Robot Data Needed?
Multimodal Data Engine: Structured signals—including electromyography, joint torques, and contact forces—are integrated into pre-training to construct a unified, hierarchical biomechanical data engine. This enables the model to learn the underlying dynamics and control intentions from human video demonstrations, incorporating dynamic signals into world modeling, energy evaluation, and motion planning. Consequently, a transferable implicit control model is formed, eliminating reliance on imitation learning based on real-robot teleoperation trajectories.
The electromyography branch serves as a prior over latent motor intent and exertion patterns. The control loop determines how to apply force, while the cognitive loop determines whether further observation or active probing is necessary. By formulating active probing as a cost-aware decision, information gain and energy-based mechanisms are unified within a single framework.
A joint loss function is defined comprising electromyography supervision error, kinematic error, and physical consistency constraints. This ensures that the data engine does not merely replicate human motions, but instead extracts fundamental principles of force generation, feedback control, and active exploration strategies from human behaviors.
The joint loss function is formulated as follows:
Self-Evolution
Self-evolution operates across three levels: online state correction, experience and skill refinement, and model and policy refinement. Attractors serve not only for goal maintenance but also as the underlying mechanism for skill memory and experience organization. Reward and energy jointly drive self-evolution:
where and denote successful and failed/suboptimal actions, respectively, and is the separation margin.
This establishes a closed loop: executionpredicting outcomesobserving actual outcomesassessing deviations and success/failureupdating experiences and modelsvalidating in novel tasks.