Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models

TL;DR

Spirit AI (千寻智能), a Beijing embodied-AI startup, open-sources Spirit v1.5 — a Vision-Language-Action (VLA) foundation model for robotic manipulation that beats Physical Intelligence’s π0.5 on the third-party RoboChallenge Table30 leaderboard (66.09 total / 50.33% success across 30 real-world tasks; first model to clear 50%). The architecture is a Qwen3-VL backbone with a DiT-based action head that takes current observations plus text and emits the next action chunk; weights, inference code, and the per-task RoboChallenge submission checkpoints are all on Hugging Face and GitHub under an open license. The blog’s title — “Clean Data Is the Enemy of Great Robot Foundation Models” — is the editorial thesis: v1.5’s gains over the team’s prior version came primarily from deliberately broadening and scaling the training distribution rather than filtering it.

Key claims

Spirit v1.5 ranks #1 on RoboChallenge’s Table30 leaderboard with a total score of 66.09 and 50.33% task success across 30 real-robot tasks, surpassing the prior leader π0.5 [Blog headline / RoboChallenge leaderboard].
Architecture: a Qwen3-VL VLM backbone provides perception + language understanding; a DiT-based “action expert” head consumes the VLM hidden states and emits the next action chunk for the robot to execute [GitHub README — model/modeling_spirit_vla.py].
Released artifacts are the full base model weights, inference code, per-task fine-tuning scripts, and the specific per-task checkpoints used for each RoboChallenge submission — explicitly framed to let the community reproduce the leaderboard result and build on top [Blog “open-source” section; HF: Spirit-AI-robotics/Spirit-v1.5].
The editorial thesis (“clean data is the enemy”) argues that aggressive demonstration filtering hurts generalization: scaling and diversifying real-world robot trajectories — including noisier, less-curated episodes — was the dominant driver of v1.5’s gains over v1.0 [Blog title + body].
Training is standard distributed PyTorch (torchrun) on a per-task fine-tuning recipe layered on top of the base model; the public training script is scripts/run_finetune.sh [GitHub README — Training section].

Method

Spirit v1.5 follows the now-standard VLA recipe — pretrained VLM as the perception/reasoning backbone, separate action-prediction head, action chunking at deployment — with two specific choices: (1) the backbone is Qwen3-VL rather than a Gemma/PaliGemma-class model, (2) the action head is a Diffusion Transformer (DiT) rather than a flow-matching MLP (π0/π0.5) or autoregressive token head (Gemini Robotics). Inputs are the current observation (images) plus a textual task description; the output is the next action chunk (default chunk size 60 in the released RoboChallenge submission scripts). The team did not publish a paper alongside the release; the blog post and GitHub README are the primary technical surface, and the headline empirical contribution is the RoboChallenge result rather than a new training algorithm.

The “clean data is the enemy” framing — which is the post’s actual title — is the team’s data-curation editorial: v1.5’s recipe deliberately keeps a broader and more diverse demonstration distribution than v1.0, on the bet that noise in real-world data is informative rather than something to filter out. The blog does not quantify this with an ablation table that’s been independently verified at filing time.

Results

RoboChallenge Table30 leaderboard: total 66.09, success rate 50.33% — first place; first model on the leaderboard to exceed 50% success across the 30 tasks [external coverage: Tencent News, Zhihu].
π0.5 (Physical Intelligence) is the displaced #1; the gap is reported on both the aggregate score and the success-rate axis [external coverage; not independently re-verified here].
Community reception: Jim Fan (NVIDIA embodied-AI lead) signal-boosted the release, Hugging Face’s official account congratulated; AdinaY’s HF post flagged it as the first Chinese-origin VLA to top the leaderboard.
No published benchmark numbers outside RoboChallenge Table30 at filing time. No ablation on the “clean data” thesis is publicly disclosed.

Why it’s interesting

This is the first filed open-weights VLA to top RoboChallenge, and it lands directly next to two other recent VLA datapoints the wiki tracks: π*0.6: a VLA That Learns From Experience (RECAP) (π0.6 + RECAP — flow-matching VLA + iterated-offline-RL on top of Gemma 3) and π0.7: A Steerable Robotic Foundation Model with Emergent Compositional Generalization (π0.7 — steerable + compositional). Spirit v1.5 is the “scale and diversify the demonstrations, ship the weights” bet, where the π line is the “improve a fixed-data VLA with on-robot RL + advantage conditioning” bet. Both claim wins on real-robot tasks; both released around the same time; they’re different answers to the same generalist-policy question.

The “clean data is the enemy” thesis also cuts against the curation-heavy framing in several filed papers — most directly A Bitter Lesson for Data Filtering (a bitter lesson for data filtering), which would predict that aggressive filtering eventually loses to scale. Spirit v1.5 is a robotics-side datapoint for that prediction. It contrasts with Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners (DVA — small curated dataset, video-as-policy), which is a deliberately data-efficient counter-bet, and complements LingBot-VLA: A Pragmatic VLA Foundation Model (LingBot-VLA — Ant Group’s 4B Qwen2.5-VL-backbone VLA on 20k hours of dual-arm data), another scale-and-open release in the same family.

On openness: along with LingBot-VLA, this strengthens the pattern that the open VLA frontier in 2026 is increasingly led by Chinese labs that ship full weights + code, while the closed flagships (Gemini Robotics 1.5 brings AI agents into the physical world, π*0.6 inference-only) keep training data and post-training recipes private.