KineWorld step-500 checkpoint

This repository contains the complete step-500.safetensors checkpoint from a KineWorld training run and its paired action_norm_stats.npz. The run completed 500 optimizer steps on 24 accelerators. This release includes a public training configuration with machine-specific paths removed.

KineWorld uses a Wan2.2-TI2V-5B video backbone with an RGB/optical-flow dual-stream model and an action-prediction module. The run used RoboTwin head-camera data, 14-dimensional actions, and robot-only optical flow. It was initialized from an earlier action/world-model checkpoint rather than trained from scratch. The warm-start checkpoint SHA-256 is 21940bd45e95dc39bc3b166c6ecd2a777efa01c8068124958778173cd01ddf66.

Files and integrity

File	Bytes	SHA-256
step-500.safetensors	13,113,498,528	86294739c54073c836a0dcb3f9114c6cf2bf83d1a8698423b71698e5f88460a3
action_norm_stats.npz	1,148	d8d5364fe6c68d2dd857ad97ef99d8ac3720de122e0ecb5acced732cc1a7d465

The Safetensors header contains 2,019 tensors. The companion action-normalization statistics are needed for the RoboTwin action-policy path. The base Wan weights, tokenizer, RoboTwin assets, and training data are separate dependencies and are not included here.

Scope

The checkpoint file is complete and its local SHA-256 was verified before upload. No benchmark score or generated-video quality has been established for this checkpoint. It has not been shown to reproduce results in the KineWorld V4 manuscript. The public code and configuration are a subset of the research implementation; some research components and the training dataset are not released. A full training reproduction is therefore not provided.

Training provenance

• Optimizer steps completed: 500.
• Training source archive SHA-256: 639dba31df48b9bf5dcd3c8ee654ec679dbb5f49ad0a74b337c59ea321a0e316. The public source subset is not a byte-identical copy of this training archive.
• Training manifest SHA-256: b21baa6a699e8972a8a3f590c676bcf1471214f4f56a2ecad5d2c9e78e302497.
• Warm-start checkpoint SHA-256: 21940bd45e95dc39bc3b166c6ecd2a777efa01c8068124958778173cd01ddf66.
• Hyperparameters and model modes: training_config_public.json.

For source-code setup and inference commands, see the KineWorld GitHub repository.