A robot foundation model built on 20,000 hours of real-world baggage

What it does LingBot-VLA is a vision-language-action foundation model for dual-arm robot manipulation. It was pre-trained on roughly 20,000 hours of real-world demonstration data across nine common dual-arm configurations, and offers both a depth-free and a depth-distilled variant. The release includes weights, training configurations, and evaluation pipelines for simulation and physical hardware.

The interesting bit The authors sell pragmatism, not just accuracy. They benchmark training throughput against existing VLA codebases and claim a 1.5× to 2.8× speedup depending on the underlying vision-language model. They also publish explicit performance tables on the GM-100 real-world benchmark and RoboTwin 2.0 simulation tasks, showing higher success rates than several recent competitors.

Key highlights

• Pre-trained on 20,000 hours of real-world dual-arm robot data across nine platforms.
• Claims 1.5×–2.8× training speedup over existing VLA-oriented codebases.
• Provides both depth-free and depth-distilled model variants.
• Ships with real-robot post-training configurations and open-loop evaluation support.
• Benchmarked on GM-100 (three physical platforms) and RoboTwin 2.0 simulation.

Caveats

• The README notes a breaking config change from LeRobot v2.1 to v3.0; older checkpoints downloaded before May 2026 need manual field removal or a re-download.
• Throughput metrics were initially mislabeled as per-GPU and later corrected to aggregate totals, so read older comparisons carefully.

Verdict Worth a look if you are building or fine-tuning policies for real dual-arm robots and care about training efficiency. Skip it if you want theoretical novelty without hardware-in-the-loop validation.