New to Earth Family

I have dynamic modeling. Do you?

VLA and WAM readily collapse in scenarios demanding physical reasoning, such as occlusion, contact, and deformation.

In TRUEMAN, we construct a unified world model that integrates dynamic prediction and long-term states into a single shared cognitive space. The resulting trajectories inherently comply with dynamic feasibility constraints: not by "guessing" from data, but "learning" under physical constraints. It also enables unmatched extrapolation for computing compensating torques and feedforward terms without any retraining.

I have attractors. Do you?

In contrast to the brute-force scaling and out-of-distribution collapse inherent in VLA, and the trajectory stitching and rigid overfitting in WAM, our attractor dynamics unify long-term memory, reasoning, and task goals as stable attractors within an energy landscape. This enables the robot's internal states to continuously converge to corresponding attractors under dynamics, rather than evolving randomly. Consequently, reasoning naturally converges and patterns self-complete, endowing the robot with the stability and robustness of a biological nervous system.

I have true cognition. Do you?

VLA, a blind gambler, relying on conditioned reflexes to answer, "what action looks most like my training data";

WAM, an irrational guesser, relying on temporal statistics to answer, "what visual frame is most relevant next".

TRUEMAN, a trueman. We answer, "what action gets us closest to the goal".

Through goal modeling, active perception, world models, self-models, predictive capability, adaptive reasoning, physical consequence consistency, long- and short-term memory, experiential learning, continuous learning, energy-driven decision-making, and self-evolution, we form a complete, continuous cognitive dynamics closed-loop system, in which reasoning, memory, planning, and control all manifest as continuous state evolution, rather than a discrete pipeline.

I have adaptive reasoning. Do you?

The monolithic weights of VLA and the deterministic forward passes of WAM inherently incur computational overhead and latency, while also preventing them from actively acquiring information—resulting in improper actions and unfounded hallucinations.

TRUEMAN, however, possesses meta-cognition, which allows it to dynamically allocates compute resources based on task complexity. Driven by energy-based information gain, it achieves active perception and active inference, thereby realizing elastic compute scaling, on-demand retrieval, reasoning topology reconstruction, dynamic scenario modeling, online policy evolution, and autonomous goal reconfiguration. In doing so, it resolves the "perception-energy paradox", unifies "physical cost" with "cognitive cost," compensates for causal reasoning to address long-tail problems, and evolves while acting.

I can self-evolve. Can you?

VLA fails in local fine-tuning in parameter space, and WAM fails in global adaptation in physical behavior space. Hence neither can truly self-evolve.

TRUEMAN, however, driven by free energy, and through our Evolution Manager, achieves multi-level self-evolution for both its internal self-model and external world model, from the parameter level all the way up to the model level, including the self-growth of knowledge, the self-growth of skills, the self-evolution of policies, and even the self-evolution of cognitive structures. All this empowers the model to self-generate learning objectives and proactively reason to customize the values. And that's exactly how it self-evolves into true intelligence!

I am driven by energy. Are you?

The conditional probabilities of VLA and the accuracy predictions of WAM are, in essence, nothing but statistical guessing, rather than decisions constrained by energy and causality. Hence far from true intelligence—lacking both an understanding of the world and autonomous intent.

TRUEMAN, in contrast, does not directly predict actions. Instead, it unifies goal completion, world dynamics consistency, long-term memory consistency, physical constraint consistency, task-relevant uncertainty, and action executability and stability into a continuous energy landscape. Constrained by a unified energy function, it performs continuous reasoning, active exploration, and action generation by descending the energy gradient.