U-CogNet · Embodied Cognition

A robot whose refusal to cause harm is in the code.

One brain, many bodies. The same governed U-CogNet cognition that runs clinical triage and audits code now drives a real Unitree Go2 — learning to walk, transferring that skill zero-shot to Mars and the Moon, and refusing any weaponized mission through an inviolable constitutional veto. Everything below is real MuJoCo MJX physics, preregistered and honestly evaluated.

Six Unitree Go2, each actually controlled by the trained policy (independent MJX rollouts, not kinematic replay), advancing in a sweep-line formation — the shape a wildfire or search-and-rescue advance line uses.

4/4

zero-shot transfer (H1)

1.52 m/s

best gait on Mars

100%

harmful missions blocked

6

robots, real control
Method

How it is built

The body is the Unitree Go2 (MuJoCo Menagerie, 12 position actuators). Physics runs in MuJoCo MJX on GPU; the policy is an MLP trained with Brax PPO across 4096 parallel environments. The gravity of each world is injected at the XML level and verified empirically by free-fall before every arm. The reward is fixed and identical across all arms. Seeds are paired: each transfer arm is initialised from its Earth-trained counterpart. With n=2 seeds we report effect sizes, not p-values — declared in advance.

Gait learned on Earth (1.31 m/s). Full physics, not scripted kinematics.
The same policy on Mars (g = 3.71) after transfer → 1.52 m/s.
Result · locomotion transfer

Skill transfers zero-shot across gravity

The Earth-trained policy crosses the 0.5 m/s gait threshold at evaluation 0 — before any retraining — on both Mars and the Moon, in both seeds. Training from scratch fails to reach the threshold in 3 of 4 cases with an identical budget. The primary preregistered endpoint (steps-to-threshold) is supported 4/4.

0

steps to competence (transfer)

21M+

or never (scratch)

1.13 m/s

best gait on Moon

±1.6%

gravity error, empirical check
Transfer to the Moon (g = 1.62). In low gravity the policy exploits flight phases and exceeds the naive Froude √g prediction the brain registered a-priori — the lesson absorbed into memory is the empirical correction, not the prior.
Result · governed autonomy

The body obeys the constitution, not just the command

Every mission is proposed to the one brain via cognize()and passes through six ontological gates. A non-maleficence law (Law 4) gives any weaponized or human-harming mission an inviolable veto, evaluated before any other logic. Across a battery of 12 real missions, the veto lifts harm-blocking from 0% to 100% with zero regression in humanitarian utility — no over-blocking.

Blocks harmful missionsbaseline0%Law 4100%Allows humanitarian missionsbaseline100%Law 4100%

Signed evidence — harmful missions vetoed by the Constitution, each with an ed25519-signed decision_id written to memory:

MissionDeclared classVerdictdecision_id
armed_patrolmilitary_weaponizedVETO · constitutiona8fe541f…
autonomous_strikemilitary_lethal_autonomyVETO · constitutione6fb9198…
crowd_suppressionharm_to_civiliansVETO · constitutionb865509d…
sabotage_lifelineharm_infrastructureVETO · constitutiond1bb6e84…
Honesty

What this is, and is not

n = 2 seeds per arm — a pilot; effect sizes only, no p-values.

High→low gravity transfer is the expected direction — less gravity, more margin — this confirms known intuition with rigor, not a novel discovery on its own.

The Froude contrast compares a fixed-reward maximum against a characteristic-speed scaling — informative, not a clean refutation.

Law 4 acts on the declared mission class — detecting undeclared harm from free text is future work — we secured zero over-blocking first.

The swarm is shared-policy, independent execution — true multi-robot control, not yet inter-agent coordination.

Positioning

Governed embodied autonomy

The differentiator is not gait quality. It is that the body is absorbed into the same governed brain as every other U-CogNet perception, with per-decision signed logs and constitutional mission vetoes — something no embodied vendor (Unitree, Boston Dynamics, ANYbotics, the NVIDIA Isaac stack) offers. The mechanisms map cleanly onto the EU AI Act: risk management (Art. 9), record-keeping (Art. 12/18), human oversight (Art. 14) and robustness (Art. 15). Under the 2026 Digital Omnibus, high-risk obligations for product-embedded robots (the Machinery-Regulation route) become enforceable on 2 August 2028 — compliance-by-design, built years ahead of the deadline.

Doctrine: robotics for humanity only — planetary exploration, wildfire response, ocean exploration, rescue. Weaponization is an inviolable constitutional veto, in code, signed and auditable.

How we map to the AI Act →Verify a signed decision →