武蔵

Musashi Labs

Close combat robotics
for the post-ballistic era

Autonomous melee weapons platforms and the reinforcement learning environments that train them. When firearms fail against armored humanoids, contact-range lethality becomes decisive.

六十

Duels fought by Musashi

零

Defeats

11.2M

Simulated engagements

零

Defeats

Scroll

01 Strategic Thesis

Firearms were designed to incapacitate
humans. Robots aren't human.

Modern small arms are optimized against a specific target: a pressurized biological system with a central nervous system. Hemorrhage, CNS disruption, hydrostatic shock — none of these apply to an armored robotic platform.

As humanoid robots enter contested environments — hardened, redundant, immune to pain and suppressive fire — the engagement model changes. A 5.56 NATO round that devastates a human torso punches a clean hole through a robot chassis and hits nothing critical.

Melee-range mechanical lethality exploits the one vulnerability that cannot be engineered away: articulating joints. Every humanoid platform, regardless of armor, exposes actuators at points of rotation. Sustained, directed force at these points achieves reliable kills that ballistic weapons cannot.

ROBOTSIM · ENGAGEMENT SIMULATION

2,400 lbs Peak shear force

14 ms Strike latency

93.7% Joint accuracy

∞ Endurance

02 Weapons Platform

One platform for contact-range lethality

ML–CQB/01 · KATA SEQUENCE

ML–CQB/01

Rōnin

Humanoid Assault Platform

Full-scale humanoid for urban close quarters. Hydraulic shear arms targeting joints. Navigates human-built environments natively. Sacrificial forward limbs protect a hardened cutting core.

Height178 cm
Mass92 kg armored
Grip2,400 lbs peak
Sprint6.2 m/s
Endurance4.5 hr ops

Distribution Limited

Anti-Personnel Applications

The Rōnin platform is designed primarily for engagement against robotic adversaries. However, the underlying capability set — high-speed limb articulation, contact-range force delivery, autonomous target discrimination — is inherently dual-applicable.

The Rōnin platform's joint-targeting system requires no doctrinal modification to engage unarmored or lightly armored biological targets. Engagement profiles against personnel are computationally simpler than robot-on-robot scenarios due to reduced target hardness and the absence of adversarial AI countermeasures. The system's training corpus includes human biomechanical models as a subset of its general engagement library.

Target discriminationCombatant / non-combatant
Engagement authorityHuman-on-the-loop
Force calibrationIncapacitate / neutralize
IHL compliance modeConfigurable ROE
Biomechanical modelsIntegrated
Acoustic signatureSub-ambient at 10 m
Visual detectionNil in low-light

All anti-personnel capabilities are subject to applicable international humanitarian law and host nation rules of engagement. Autonomous lethal engagement requires explicit authorization per DoD Directive 3000.09. Platform supports full audit logging of all engagement decisions.

03 Dojo RL Environment

For a thousand years
Rōnin has studied the blade

Musashi wrote that the warrior must know every art, so we trained a multimodal foundation model on poetry, art, military philosophy, and a thousand years of simulated combat.

At 10,000× real-time, every Rōnin accumulates a millennium of blade study before its first real-world engagement. Emergent technique discovered through self-play against increasingly hardened adversaries — informed by the same aesthetic intuition that governs a brushstroke or a line of verse. Combat as art. Art as combat.

Run one-two-four-six— one thousand autumn leaves fall. None reach the ground whole.

— Rōnin v0.3, epoch 1,246, loss threshold achieved

i

MuJoCo-Extended Physics

Contact-rich simulation with deformable joint models and armor penetration dynamics
ii

Adversarial Self-Play

Population-based training across 2,048 concurrent agents discovering novel technique
iii

Sim-to-Real Transfer

Domain randomization across 140+ parameters. Zero-shot hardware deployment, <4% gap
iv

Procedural Urban Terrain

Hallways, stairwells, rubble fields. Agents learn to use structure as weapon

dojo-env v3.2 — training session

04 Design Doctrine

First principles

一

Joints are the universal weakness

Every articulating platform exposes actuators at points of rotation. We build weapons that exploit this immutable constraint of humanoid geometry.

二

Disproportionate force when outnumbered

A single Rōnin engages multiple adversaries simultaneously. Superhuman reaction latency, 360° threat awareness, and the absence of self-preservation instinct mean that outnumbering the platform does not produce the expected result.

三

Emergent technique, not choreography

We don't program combat. We build environments where effective technique is discovered through billions of adversarial encounters in simulation.

四

The factory is the weapon

Decisive advantage belongs to whoever manufactures and deploys faster. Our platforms are designed for production at scale — not bespoke engineering.

Close combat roboticsfor the post-ballistic era

Firearms were designed to incapacitatehumans. Robots aren't human.