The causal
transformer.
we have deprecated rigid state-space models. our neural kernels utilize causal self-attention over historical proprioceptive state and multi-modal sensory embeddings to generate deterministic control torques at 500hz.
SYSTEM_CAPABILITIES_01 // TRANSFER_DYNAMICS
Closing the reality gap.
a policy that only works in simulation is a toy. the large behavior model uses extreme domain randomization (edr) and causal context to infer reality.
System Identification
by attending to the kv-cache of past actions and their physical results, the transformer naturally infers unmodeled parameters like exact payload mass or floor friction.
Domain Randomization
during parallel gpu training, mass matrices, motor latency, and sensor noise are aggressively randomized. the policy learns invariant representations of physics.
Zero-Shot Deployment
because the latent space is completely regularized against noise, we achieve 0.94+ sim-to-real correlation without requiring any physical fine-tuning.
SYSTEM_PIPELINE // EXECUTION_LEDGER_v4
Tokenizing physics.
a deterministic four-stage process for bridging raw kinematic vectors to the high-dimensional latent space of the causal transformer core.
Continuous State Encoding
raw telemetry (joint angles, velocities, imu vectors) concatenated.
Latent Token Embedding
linear projection into hidden dimension with temporal encodings.
Causal Self-Attention
dot-product over past 8,192 Keys in KV-Cache for historical context.
Torque Decoding
mlp head decoding latent tokens back to physical actuator targets.
SYSTEM_CAPABILITIES_02 // EXECUTION_FILTER
Deterministic safety.
neural outputs are inherently non-deterministic. to prevent hardware damage, every action vector is passed through our kinematic supervisor—an unyielding c++ layer enforcing constraints.
if limits exceeded, trigger high-damping pd stance loop to arrest momentum smoothly.
STATE: PD_STANCE_READYSYSTEM_CAPABILITIES_03 // EDGE_TOPOLOGY
Direct bus interfacing.
the output of the kinematic supervisor is piped directly into real-time industrial communication buses. we bypass standard os kernels to maintain strict microsecond precision required for dynamic locomotion.
For high-performance motor drivers (e.g. Elmo, Moog).
For distributed lower-power actuator networks.
SYSTEM_READY
Initialize execution sequence.
the private beta is active. allocate compute nodes, synthesize your reward landscapes, and export zero-shot neural policies to edge hardware.