HelixDualSystemRunner<T>

Models & Types · AiDotNet.VisionLanguage.Robotics

Streaming controller that coordinates Helix’s fast System-1 visuomotor policy (200 Hz) and slow System-2 VLM (7–9 Hz) per Figure AI’s dual-rate dual-system architecture.

How It Works

Per Helix (Figure AI 2025, arXiv:2502.07092), high-level intent runs slowly while low-level control runs fast. This runner is the explicit coordinator: each call to Step advances one S1 tick and lazily re-invokes S2 when the cached latent has gone stale (every System2TicksValid S1 ticks — default 22 to match S1:S2 = 200Hz : ~9Hz).

The runner is generic over the model’s S2 and S1 callbacks so the same coordination logic is reusable: Helix wires its native VLM into system2Forward and its action head into system1Forward. Future Helix-class models (e.g. Figure’s later iterations) can reuse this runner unchanged.

Constructors

Constructor Summary
HelixDualSystemRunner(Func<Tensor<T>, string, Tensor<T>> system2Forward, Func<Tensor<T>, Tensor<T>, Tensor<T>> system1Forward, int system2TicksValid) Constructs a runner.

Properties

Property Summary
HelixSystem2Latent<T> CachedLatent { get; } The currently cached S2 latent, or null if no S1 step has been taken yet.
int CurrentTick { get; } Current S1 tick counter since the runner was created or Reset was called.
int System2TicksValid { get; } Number of S1 ticks an S2 latent remains valid (default 22 — paper §4.1: S1 @ 200 Hz, S2 @ 7–9 Hz).

Methods

Method Summary
Tensor<T> Step(Tensor<T> observation, string instruction) Advances one S1 tick.
Tensor<T>[] Rollout(Tensor<T> observation, string instruction, int numSteps) Runs numSteps consecutive S1 ticks against the same observation/instruction (e.g.
void Reset() Resets the tick counter and invalidates the cached S2 latent.