Spirit v1.5

Spirit AI · 2026-01

VLANew in 20265BNot disclosed

Trained on diverse, loosely controlled collected data instead of curated data.

Paper

No paper

Backbone LLM: Qwen3 · Qwen3-VL-4B config
Vision encoder: Qwen3-VL ViT
Action head: DiT action head (latent action tokens, secondary) (Not disclosed)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): Qwen3-VL-4B-Instruct VLM + DiT action head (latent action token prediction, secondary)

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed (README recommends 8× A100 80GB)
Data type: Real-robot teleoperation (diverse·loosely controlled collection)
Data scale: Not disclosed (hours·episodes)
Main datasets: Not disclosed
Tasks & evaluation: RoboChallenge Table30 (30 tasks) 1st place; overall 66.09, success rate 50.33% (secondary)
Robot: Not disclosed

Reinforcement learning: Not disclosed · Training method not disclosed
Base control: Fixed · No base control (RoboChallenge Table30 tabletop manipulation; explicitly excludes non-manipulation scenarios)

Key features: Trained on diverse, loosely controlled collected data instead of curated data. Ranked 1st on RoboChallenge Table30 as of 2026-01-11.
Notes: License conflict: HF apache-2.0, GitHub LICENSE MIT. README '[2026-01] Initial Release'; date 2026-01-12 is secondary. Effect of collection method (secondary): effective collection time per person up about 200%, expert input down about 60%, fine-tuning iterations down about 40% at the same pretraining scale.
Verification: Confirmed