Learning & adaptation
Custom RL environments, sequential models, behavioral priors, and task-specific refinement.
RL / ML / LATENT REPRESENTATIONSMachine learning · Reinforcement learning · Embodied AI
Intelligence is
more than
a model.
We build the systems around it. From noisy data to adaptive behavior, our work connects learning, simulation, and real-world execution.
THE SYSTEM AROUND THE MODELCM / 01
INPUTNoisy environmentsSequential observations · Uncertainty
LEARNING & CONTROLRepresent.
Adapt. Act.Motion priors + Task intelligence
Adapt. Act.Motion priors + Task intelligence
VALIDATIONMeasured responseSimulation · Inference · Execution
OBSERVE → LEARN → VERIFY
From quantitative research to embodied intelligence.Discover our approach
The focus
Research that connects
learning to execution.
Models are one layer. Data quality, objective design, system integration, and validation determine what survives outside the training environment.
Simulation & validation
Cross-simulator evaluation, domain randomization, and tests designed to expose hidden assumptions.
ISAACLAB / MUJOCO / ROBUSTNESSIntegrated deployment
Aligned observations, calibrated actions, deterministic inference, and end-to-end instrumentation.
ONNX / CONTROL / INSTRUMENTATIONCurrent workstream
From motion knowledge
to physical behavior.
Our current research adapts NVIDIA’s Generative Pre-trained Controller architecture to a Unitree G1 deployment stack.
01/ 04
Build a useful motion vocabulary.
Finite scalar quantization, discrete latent representations, joint-order canonicalization, and frame alignment.
Encoder / Quantizer / Decoder02/ 04
Learn a task without losing the prior.
Proprioceptive conditioning, supervised adapter initialization, and parameter-efficient reinforcement-learning refinement.
Prior / Embodiment / Task03/ 04
Challenge the assumptions of simulation.
IsaacLab and MuJoCo checks, domain randomization, observation parity, and perturbation testing.
Noise / Contact / Latency04/ 04
Preserve correctness through inference.
ONNX export, history buffers, output scaling, scheduling, and synchronization with low-level actuation.
Observe / Infer / ActCurrent robotics research architecture · select a stage to explore
Embodied AIActive research
Generative control for a physical embodiment.
Adapting pretrained motion structure to the joint topology, observations, and control cycle of the Unitree G1.
Explore the workSystemsActive research
Closing the simulation-to-execution gap.
Testing what survives a change in simulator, operating conditions, and inference runtime.
Explore the workSequential learningResearch foundation
Learning from noisy market environments.
A quantitative research stack that shaped our approach to uncertainty, reward design, and honest evaluation.
Explore the workOur conviction
More model complexity
does not automatically
create more signal.
Our first proving ground was financial markets. It taught us to separate model performance from system correctness—and to test both.
Read our storyStart a conversation
Bring us a hard
learning problem.
Discuss a research collaboration, explore a training pipeline, or request a walkthrough of our work.
Discuss your project