Products & Services

Robot logs, sensor data, human operation histories, task videos, success and failure records, and simulation results are transformed into data that Robot Decision Model can learn from.

We are researching workflows that ingest hardware data from recording environments such as LeRobot and simulation results from tools such as NVIDIA Isaac Sim, organizing them into State / Action / Outcome / Preference.
Beyond storing robot data, we build a structure centered on decisions.

• What was the current state?
• Which candidate actions were available?
• Which action was selected?
• What was the outcome?
• Which action was more appropriate than the alternatives?

Using Robot Decision Dataset, we develop dedicated models to evaluate and select the next action for robots and humanoids.

Instead of calling a large frontier model over a network from the robot, we transfer decision-making capabilities from powerful Teacher Models and simulation environments into smaller models through distillation and fine-tuning.

• Action candidate evaluation
• Success / Risk / Cost prediction
• Pairwise action comparison
• Skill selection
• Uncertainty estimation
• Recovery decisions

The future model hierarchy is envisioned to range from an 8B-class general-purpose Robot Decision Model to very small Decision Models for specific processes.

Instead of using the same model, candidate count, and compute for every situation, decision paths change according to the robot’s objectives, data, risk, and current state.

Task-specific Harness Profiles are built on a fixed execution kernel, combining multiple decision patterns.

Direct / Fast

Use a lightweight model for immediate decisions in clear states.

Verified Selection

Generate multiple actions and compare them with a Verifier.

Predictive Simulation

Simulate candidate actions in high-risk situations.

Recovery

Reassess and replan after failures or in unfamiliar states.

Adaptive Decision Compute scales the Decision Model, candidate count, Verifier strength, simulation depth, and retry count up or down at runtime.