Applied Intuition · Research

Scaled Simulation, Minimal Demonstration

Terra series research created by Applied Intuition aims at pushing the boundary of physical AI learned in closed loop with minimal human demonstration. Terra series exploits high-throughput, reactive simulation at scale together with efficient closed-loop reinforcement learning to bring the generalizability and robustness of physical AI to the next level.

TerraZero Technical Report Developing state-of-the-art autonomous planners and simulation agents through self-play: powered by a reactive, high-throughput simulator executing millions of steps per second, alongside a procedural engine that generates an endless supply of edge cases — all at a minimal cost. View → TerraTransfer Teaching an E2E planner by the highly generalizable and robust planner from self-play, transferring the competence of TerraZero policy into E2E with video input without imitating human demonstration. View → SPACeR ICLR 2026 Human-like, reactive simulation agents via self-play anchored to a reference model — cheap, realistic traffic to train against. View →

TerraZero · Technical Report

Procedural Simulation and Self-Play at Scale

A fast object-level simulator and a procedural scenario engine that treat logged data as a starting point, not a ceiling — bootstrapping real maps and logs into an effectively unlimited supply of low-cost, high-value scenarios for agents to learn from. A compute-efficient self-play recipe then learns from that stream through trial and error, without human demonstrations or imitation of logged trajectories. The resulting policies reach state-of-the-art closed-loop performance and transfer zero-shot across cities and datasets.

Project Page → Technical Report (PDF) arXiv (on release)

TerraTransfer

End-to-end learning via self-play without expert demonstrations

TerraTransfer uses TerraZero's self-play planner as a teacher and distills its competence into an end-to-end camera policy — no human demonstrations required. It aligns the vision policy's latent space to the planner's vector representation, achieving 0.490 HD-Score on HUGSim and outperforming imitation baselines by +0.130.

Project Page → arXiv

SPACeR · ICLR 2026

Human-Like Simulation Agents via Self-Play Anchoring

SPACeR anchors TerraZero's self-play to a pretrained tokenized reference model, producing sim agents that are simultaneously human-like and reactive — low-cost, realistic traffic to train and evaluate against. The policy is ~50× smaller than tokenized baselines with 10× faster inference, and outperforms prior self-play and imitation methods on the Waymo Sim Agents Challenge.

Project Page → arXiv Paper

Join Us

Building physical AI that moves the world.

The Terra Series is developed by the research team at Applied Intuition. We are hiring researchers and engineers to push autonomy from simulation into the real world.

Careers at Applied Intuition Research