SPRITE PENDING drop image at
/sprites/<file> ROUND·01
ADVERSARIAL VLA
robustness of vision-language-action systems under live adversaries
- ROBUST
- INTERP
- SCALE
- ADV
WHAT IS Φ ?
Φ is an independent research collective for physical-world AI fighting. The symbol is Φ. We pronounce it "fight".
We study how physical-world AI systems — vision-language-action models, world models, and whatever architectural categories come next — learn, fail, and adapt when made to fight other physical-world AI systems. The fighting we care about is what happens inside the intelligence: the representations, the world models, the failure modes, the strategic emergence when one physical AI is set against another.
We are not a company. We are not a lab in any institution. We are a small group of independent researchers working at the intersection of multi-agent RL, VLA & world models, mechanistic interpretability, contact-rich dynamics, and adversarial self-play. Φ is a ten-year project. We publish openly.
- FIELDadversarial multi-agent embodied AI
- MODEindependent · open · part-time · distributed
- FOUNDERLiu Yuchen — HKUST
- EST.2026 · ten-year arc
CHOOSE YOUR FIGHTER
active and planned research vectors
SPRITE PENDING drop image at
/sprites/<file> ROUND·02
CONTESTED WORLD
world models when an opponent shapes the dynamics
- ROBUST
- INTERP
- SCALE
- ADV
SPRITE PENDING drop image at
/sprites/<file> ROUND·03
SELF-PLAY CIRCUITS
mechanistic interpretability of self-play policies
- ROBUST
- INTERP
- SCALE
- ADV
SPRITE PENDING drop image at
/sprites/<file> ROUND·04
ENERGY-BOUNDED
strategic emergence under hard physical constraints
- ROBUST
- INTERP
- SCALE
- ADV
SPRITE PENDING drop image at
/sprites/<file> ROUND·05
SAMPLE-EFFICIENT
opponent-distribution coverage without league-scale compute
- ROBUST
- INTERP
- SCALE
- ADV
SPRITE PENDING drop image at
/sprites/<file> ROUND·06
CROSS-EMBODIMENT
adversarial generalization across body morphologies
- ROBUST
- INTERP
- SCALE
- ADV
FIGHTERS
the people who do the work
PORTRAIT PENDING
PLAYER · 01
LIU YUCHEN
FOUNDER · researcher
HKUST · undergraduate
Started Φ in 2026. Working on world models under adversarial dynamics and the mechanistic interpretability of self-play policies.
@liu.yuchen
PORTRAIT PENDING
PLAYER · 02
CHALLENGER · 02
co-founder · pending
HKUST
Recruiting. If you are doing serious research in physical-world AI fighting and want a long-term home for it, the slot is open.
@???
FIGHT RECORD
publications & status updates · open access by default
- 2026 · 05 STAGE 0 Φ founded · domain online · first grant applications submitted —
- 2026 · Q4 EXPECTED first preprints · ICLR 2027 submissions arXiv
- 2027+ PIPELINE sustained publication output across ICLR / NeurIPS / ICML / CoRL / RSS open access
No preprints yet. First arXiv submissions expected late 2026 / early 2027. All Φ-affiliated work appears under open licenses with reproducible code, on arXiv and at appropriate venues (ICLR, NeurIPS, ICML, CoRL, RSS, and their workshops).
WHY THIS REGIME
Mainstream physical AI optimizes single agents in cooperative or static environments. Pick up the cube. Walk on the terrain. Don't fall. This is not the regime in which physical intelligence will actually have to operate.
The hardest, most informative tests of a physical AI system are not how well it performs alone, but how it behaves when another physical AI is actively trying to defeat it. This is where:
- world models break in the most interesting ways
- self-play produces the most surprising emergent behaviors
- VLA systems reveal their hidden vulnerabilities
- interpretability has the most leverage
- realistic curricula for general physical intelligence will eventually come from
The regime sits awkwardly between multi-agent RL (mostly toy domains), embodied AI (mostly single agents), and interpretability (mostly language models). Φ exists to occupy that intersection seriously.
WHAT WE ARE NOT
- NOTa startup, no equity, no product roadmap
- NOTa graduate program or substitute for one
- NOTa credential — Φ work makes a Φ researcher, not the badge
- NOTa brand for status
- NOTa robotics fighting league — the fight is inside the AI
CHALLENGE US
serious researchers in adversarial physical AI, self-play interpretability, or contested world models — reach out.
GITHUB github.com/phi-monster MAIL hello@φ.monster
STATUS Stage 0 · first preprints late 2026 / early 2027