Multimodal Synthesis Studio · Kokechan × ReplaySims × Ai2

Saccadic Budget: 114ms (<150ms) Awning: 1,276 Triangles Molmo Grounding: Active

Ai2

Molmo Point Grounding [u: 0.524, v: 0.612] Inclusive Copy ✓

Click anywhere on the Kyojima canvas to test Molmo pointing. Currently focused on Kyojima Eki with its preserved 1,276-triangle awning, shoji lattice, and washi banners.

01. Kyojima Eki (Exhibition Anchor)

16,702 close tris · Awning intact

02. KocoGarden / Hatsune Wing

1,360 close tris · Courtyard flora

03. Koganecho Studio Lattice

Artisan viaduct spaces · 42 places

Robot: Unitree G1 (29-DoF) Physics: 500 Hz Symplectic Euler

Timeline: Frame 0 / 180 (0.00s)

Base Tilt: 0.4°

Jitter: 0.09 rad/s²

Left Ankle: 12.4 N·m

Loss Receipt: SHA-256 #8f2a

🔬 MolmoSpaces & MolmoBot Manipulation Bridge Zero-Install MuJoCo WebGL Pipeline

Directly mapping Ai2's 42 million physics grasp annotations and 230,000 MolmoSpaces into ReplaySims without requiring 50 GB of NVIDIA Isaac Sim dependencies. Policy evaluations stream into browser VRAM with full telemetry receipts.

Foundational Navigation Model Benchmark

Embodiment-Agnostic World Model (KRM Mono) vs. Semantic Navigator (Ai2 RING)

KRM Mono (Kanaria Robotic Model)

• Input: Ego-centric video frame sequence
• Encoders: Image (spatial), Video (motion), Geometry (depth/pose)
• Token Pool: 20k tokens compressed via cross-attention into vector $z$
• Output: Next $N$ poses trajectory + residual RL policy

Ai2 RING (Robotic Indoor Navigation Generalist)

• Input: Semantic room goal + topological graph memory
• Architecture: Transformer policy trained across 230k simulation environments
• Core Strength: High zero-shot transfer across distinct robot chassis
• Output: Waypoint navigation & affordance obstacle clearance