LingBot-VLA: A Pragmatic VLA Foundation Model

TL;DR

LingBot-VLA is a 4B-parameter open Vision-Language-Action foundation model from Ant Group, built on Qwen2.5-VL-3B-Instruct and trained on 20,000 hours of real-world dual-arm robot data spanning nine popular robot configurations. It is released in two variants (with and without a depth-distillation auxiliary), claims state-of-the-art real-world success rates against WALL-OSS, GR00T N1.6, and π₀.₅ on the GM-100 benchmark (3 robot platforms), and ships a custom training stack that the authors claim achieves a 1.5–2.8× training-throughput speedup over existing VLA codebases. The release is open: weights, code, post-training recipes, and the GM-100 benchmark dataset are all on Hugging Face / ModelScope under Apache-2.0.

Key claims

20,000 hours of real-world dual-arm robot pre-training data covering nine robot configurations is used; this is positioned as the largest pretraining corpus for an open VLA [§Pretraining Data].
On the GM-100 benchmark across three platforms (AgiBot G1, AgileX, Galaxea R1Pro), LingBot-VLA with depth attains 17.30% average SR / 35.41% average PS, beating π₀.₅ (13.02% / 27.65%), GR00T N1.6 (7.59% / 15.99%), and WALL-OSS (4.05% / 10.35%) [GM-100 table].
On RoboTwin 2.0 simulation, LingBot-VLA-w/-depth reaches 88.56% Clean / 86.68% Randomized average SR vs π₀.₅ at 82.74% / 76.76% [RoboTwin table].
A depth-distilled variant (LingBot-VLA-4B-Depth, trained against the MoGe-2 / LingBot-Depth signal) consistently improves over the depth-free variant on both real-world and sim benchmarks [GM-100 + RoboTwin tables].
The training stack is 1.5–2.8× faster than existing VLA codebases at matched configurations, measured on Qwen2.5-VL-3B-π and PaliGemma-3B-pt-224-π across 8–256 GPUs [§Efficiency].
Co-designed with Baidu Cloud, the training stack reduces GPU memory consumption by 29.2% [§Acknowledgement].

Method

LingBot-VLA is a π-style VLA architecture: a frozen-or-finetuned VLM backbone (Qwen2.5-VL-3B-Instruct is the released base; PaliGemma-3B is also benchmarked) produces visual + language features that condition an action expert. The released variant follows the π₀ family’s action-expert structure, with action chunks of length 50 for sim / 25 for real-robot deployment. A depth-distilled variant additionally injects a depth signal from MoGe-2-vitb-normal and a “LingBot-Depth” pretrained checkpoint at training time, which lifts both Clean and Randomized SR on every platform. The codebase is built on top of LeRobot v3.0 with VeOmni-style training infrastructure and Baidu Cloud memory optimizations; inference supports Torch compile and a tunable number of denoising steps for the action diffusion head (default 5 for real-robot deployment speedup). Post-training is the intended downstream workflow: users prepare LeRobot v3.0 datasets, write a robot-config YAML mapping their state / action / image features, compute normalization statistics, then fine-tune one of the released checkpoints with or without depth on their robot.

Results

Headline numbers, all from the README’s reported tables:

GM-100 real-world, 3 platforms, average: LingBot w/ depth 17.30% SR / 35.41% PS; LingBot w/o depth 15.74% / 33.69%; π₀.₅ 13.02% / 27.65%; GR00T N1.6 7.59% / 15.99%; WALL-OSS 4.05% / 10.35%.
GM-100 per-platform best: Galaxea R1Pro 20.98% SR (w/ depth); AgileX 18.93% SR (w/ depth); AgiBot G1 12.82% SR (w/o depth) — the depth-free variant actually wins on AgiBot G1 SR, hinting depth-distillation isn’t strictly dominant across platforms.
RoboTwin 2.0 sim, 5 tasks, average SR: LingBot w/ depth 88.56% Clean / 86.68% Randomized; LingBot w/o depth 86.50% / 85.34%; π₀.₅ 82.74% / 76.76%. The Randomized–Clean gap is much smaller for LingBot (≈2 pts) than for π₀.₅ (≈6 pts), suggesting better generalization to scene perturbation.
Training throughput: 1.5–2.8× speedup over baseline VLA codebases for both Qwen2.5-VL-3B-π and PaliGemma-3B-pt-224-π across 8 / 16 / 32 / 128 / 256 GPU configurations; updated throughput values are reported as total aggregate (the README explicitly corrects an earlier per-GPU mislabeling).
Memory: 29.2% reduction in GPU memory consumption from the Baidu Cloud joint-optimization work.

No quantitative comparison against the recent NVIDIA Isaac GR00T N1.6 is shown beyond the GM-100 SR numbers; no comparison against OpenVLA, RT-2, or Octo is reported.

Why it’s interesting

This is the second open large-scale VLA release filed this week, after NVIDIA’s Isaac GR00T N1.6 (NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry), and the first one from a non-US lab that openly publishes both the pretraining-scale dataset hours, the GM-100 benchmark, and a training-throughput-optimized codebase. The depth-distillation design echoes the same “world-model-as-prior” pattern NVIDIA uses with Cosmos Reason inside GR00T — here it’s MoGe-2 + LingBot-Depth as a geometry prior rather than a video world model, but the structural move is similar. The Randomized/Clean gap on RoboTwin is the most interesting empirical signal: a 2-point degradation under scene randomization vs π₀.₅’s 6 points suggests the depth signal (and possibly the 20K-hour real-world pretraining mix) is buying genuine visual-distribution robustness, not just in-distribution accuracy.