KRONOS
- KRONOS-1M keeps one unbroken session across terrain, 60 Hz motion, LiDAR, and operator rubrics inside a 1,000,000-token context.
Core trick
On each failed 0.5-second step, a sandbox writes four counterfactual rollouts, one chosen and three rejected, as DPO preference pairs.
Description
- KRONOS-1M keeps one unbroken session across terrain, 60 Hz motion, LiDAR, and operator rubrics inside a 1,000,000-token context.
- It splits continuous robot trajectories into timed sub-tasks and writes past-and-future action captions for each segment.
- It scores the whole episode, then labels every 0.5 second step as normal, suboptimal, or failure.
- It grades trajectories against natural-language rubrics, and raises its thinking level when a step fails.
- On a failure it checks physics in a sandbox and writes four counterfactual rollouts (one chosen, three rejected), including physics and robot-shape changes, as preference-training pairs.
README excerpt
# KRONOS-1M // Gemini 4 Argon Managed Agents & Multiverse DPO Robotics Data Engine **KRONOS-1M** is a robotics simulation and data-engine prototype powered by Gemini 4 Argon (`gemini-4-argon`), Gemini Managed Agents (`antigravity-preview-09-2026`), a browser 3D Rapier physics arena, and PyTorch CUDA DPO policy training on a local NVIDIA GeForce RTX 3080 Laptop GPU (`cuda:0`). ## What it does - **3D Sim2Real Arena (React + Three.js + Rapier3D):** Streams procedural terrain sectors, 32-beam LiDAR telemetry, wheel-slip ratios, and 0.5s process reward ticks. - **4-Branch Counterfactual Multiverse DPO Forking:** When a simulated 0.5s failure step occurs, KRONOS forks 4 counterfactual branches (`Alpha, Beta, Gamma, Omega`) and exports Chosen vs. Rejected DPO preference pairs (`runs/brev/kronos_dpo_dataset.json`), MuJoCo MJCF XML (`runs/brev/kronos_world.xml`), and OpenUSD ASCII stages. - **Multimodal Raw `.mp4` Video Evaluation:** Sends raw `.mp4` bytes (`inline_data: video/mp4`) to `gemini-4-argon` (`thinking_level: "high"`) for 1–5 outcome reward grading with visual evidence. - **Local RTX 3080 CUDA DPO Policy Training:** Trains a PyTorch `GaussianPolicyNet` via contrastive DPO (`server/gpuDpoTrainer.py`) on `cuda:0` and saves weights to `runs/brev/kronos_policy_dpo.pt` and metrics to `runs/brev/dpo_training_metrics.json`. --- ## Table 1: Published NVIDIA FLORA Paper Baselines (Oct 20, 2026 Reference) For context, below are the published baseline numbers from *"FLORA: An Op