MODEL_ZOO // REGISTRY_v4.2

Neural weights for
physical agents.

access foundational large behavior models (lbms) trained inside our hyper-parallel simulation engine. zero-shot ready for edge deployment.

SYSTEM_MODELS // WEIGHT_MANIFEST

01
LBM-H-v4.0

Model-01 Humanoid Base

foundation humanoid policy optimized for high-dimensional coordination and bipedal stability. utilizes causal transformer core for historical state awareness.

REF_TOPOLOGY: HUMANOID
DYN_MATRIX: SOLVED // DLS: ACTIVE
WEIGHTS1.2B
CTX_WINDOW8192_STEPS
SIZE4.2 GB
SUCCESS_RATE94.2%
LATENCY1.2ms (ORIN)
SIM_TO_REAL0.89
PULL_COMMAND
iacon pull LBM-H-v4.0
02
LBM-A-v3.2

Model-02 Industrial Arm

high-precision manipulation kernel for 7-DOF industrial arms. zero-shot transfer from humanoid latent embeddings achieved in v3.2.

REF_TOPOLOGY: MANIPULATOR
DYN_MATRIX: SOLVED // DLS: ACTIVE
WEIGHTS310M
CTX_WINDOW2048_STEPS
SIZE1.2 GB
SUCCESS_RATE99.8%
LATENCY0.4ms (ORIN)
SIM_TO_REAL0.97
PULL_COMMAND
iacon pull LBM-A-v3.2
03
LBM-B-v2.1

Model-03 Bipedal Agile

specialized policy for high-impact bipedal activities. trained via extreme domain randomization (edr) for robust jumping and sprinting.

REF_TOPOLOGY: BIPED
DYN_MATRIX: SOLVED // DLS: ACTIVE
WEIGHTS680M
CTX_WINDOW6144_STEPS
SIZE2.1 GB
SUCCESS_RATE91.5%
LATENCY0.8ms (ORIN)
SIM_TO_REAL0.86
PULL_COMMAND
iacon pull LBM-B-v2.1
04
LBM-Q-v3.8

Model-04 Quadruped Base

robust quadrupedal locomotion kernel. designed for rough-terrain navigation and high-speed dynamic recovery.

REF_TOPOLOGY: QUADRUPED
DYN_MATRIX: SOLVED // DLS: ACTIVE
WEIGHTS450M
CTX_WINDOW4096_STEPS
SIZE1.8 GB
SUCCESS_RATE98.1%
LATENCY0.6ms (ORIN)
SIM_TO_REAL0.94
PULL_COMMAND
iacon pull LBM-Q-v3.8
AUTH:ENCRYPTED
ENGINE:IACON_CORE_v4
TOTAL_REGISTRY_SIZE: 16.8 GB // 0x4F2A_MANIFEST

Safety_Architecture

all policies execute behind a deterministic kinematic supervisor. if a neural command violates physical joint limits or balance constraints, a high-damping pd fallback instantly arrests momentum.

Sim_to_Real_Gap

we achieve 0.94+ correlation using extreme domain randomization (edr). by varying mass matrices, friction, and latency during parallel training, policies learn invariant representations of physics.

Simulation_Kernel

our engine resolves closed-loop kinematic chains at 10,000 tps using damped least squares (dls) solvers, eliminating mass-matrix singularities and preventing reward exploitation during rollout.

DEPLOYMENT_READY

Flash to edge silicon.

pull these weights directly via the iacon cli. our compilation engine handles the int8 quantization and tensorrt graph generation automatically.

NVIDIA_JETSON
TENSORRT_EXPORT
INT8_QUANTIZED