3 Billion Expert Transitions Across 11 Million Physics-Based Tasks (Kinetix-Offline)

TL;DR

FLAIR (Oxford) releases Kinetix-Offline, a 326 GB dataset of ~3 billion expert transitions across 11M unique procedurally-generated 2D physics tasks in the Kinetix environment. Trajectories are collected by training a separate specialist PPO agent per level and saving only successful rollouts; raw environment state is stored (not pre-rendered frames), so downstream training can render into pixel or symbolic observations on the fly and evaluate online from any point in a trajectory. As a proof of concept, behaviour cloning on this data followed by PPO fine-tuning matches or beats from-scratch PPO at a fraction of the compute — under 24 GPU-hours of BC training already outperforms the from-scratch baseline from the original Kinetix paper. The bet is that 10M+-task offline datasets unlock research into how offline agent performance scales with task diversity, world-model training, and level generation.

Key claims

The dataset covers 11M unique Kinetix levels and ~3B transitions, ~1000× more tasks than XLand-100B’s ~30k [Overview, Table]. Five splits vary by env size (small/medium/large) and expert training budget (1M or 10M steps).
Every trajectory comes from a level-specialist PPO agent, and only successful rollouts are retained [Methodology §2]. This is a per-task self-play data-collection procedure, distinct from single-generalist-policy rollout corpora.
Storing raw environment state (not rendered frames) is what makes 3B transitions fit in 326 GB and lets the renderer be swapped at load time — same data trains pixel or symbolic agents [Methodology §2, Data-loading example].
BC on Kinetix-Offline is a compute-efficient warm-start for online PPO fine-tuning: <24 GPU-hours of BC already outperforms the from-scratch PPO baseline from the original Kinetix paper (arXiv:2410.23208) [Results §5]. Not competitive with the latest Kinetix state of the art (arXiv:2603.06009), but a strong bootstrap.
Evaluation protocol splits into (a) training levels, (b) held-out same-distribution levels, and (c) hand-designed levels — enabled by the raw-state storage, which lets any trajectory point serve as an evaluation initial state [Evaluation §4].

Method

Kinetix is a JAX-based procedurally generated 2D rigid-body physics environment; each task is a level with dynamic scene objects where the goal is to bring green and blue objects into contact without green touching red. For every level in an 11M-level curriculum, the authors train a dedicated PPO specialist agent for either 1M or 10M environment steps and save only the trajectories that solve the level. Trajectories are stored as raw environment state (position, velocity, contact) rather than pixels, so the observation modality (pixel or symbolic) is chosen at load time by the downstream user via the released TrajectoryDatasetManager. The reference offline pipeline is standard behaviour cloning on the successful trajectories (experiments/offline_bc.py), optionally followed by online PPO fine-tuning that queries the same simulator.

Results

Dataset scale: 11M unique levels, ~3B transitions, 326 GB total on disk across five splits (1M/s: 6M levels / 1.5B transitions / 123 GB; 1M/m: 3.5M / 884M / 98 GB; 1M/l: 1M / 268M / 82 GB; 10M/s: 637k / 163M / 12 GB; 10M/m: 422k / 108M / 11 GB) [Overview table].
BC warm-start: <24 GPU-hours to reach a policy that outperforms the original-paper from-scratch PPO baseline [Results §5]; strong but not state-of-the-art versus [arXiv:2603.06009].
Released alongside data loaders in the main Kinetix repo with an example (examples/example_data_loading.py) and the full BC training script.

Why it’s interesting

Slots into the concept space at the intersection of two active clusters: it’s an RL environment platform release (like Genesis World — Simulation platform for general-purpose robotics & embodied AI learning and SETA: Scaling Environments for Terminal Agents (env repo)) whose distinguishing move is shipping the offline data substrate — not just the sim — at a scale (11M tasks) that dominates prior offline-RL corpora by 2-3 orders of magnitude. The specialist-per-level data-collection recipe is a strong instance of self-play-driven Synthetic Training Data in the RL setting: no human demonstrations, no LLM curation gate, just per-task specialists filtered by success. And the “store raw state, render at load time” design mirrors Worlds in One Demo: A Synthetic Data Engine for Learning Open-World Mobile Manipulation WANDA’s “reconstruct once, recompose many times” logic — the substrate is deliberately factored to serve multiple downstream training regimes. Open question the release explicitly poses: does offline agent performance scale predictably with task diversity into the 10M-task regime, and does this dataset become a testbed for level generators and world models the way ImageNet became one for classifiers.