Source note
StarVLA- α α : Reducing Complexity in Vision-Language-Action Systems
Summary
StarVLA-α argues that a plain VLM-to-action setup is enough to reach top-tier VLA performance across several robot benchmarks. The paper’s main claim is methodological: many common additions in VLA systems change results less than backbone choice and clean training setup.
Problem
- VLA papers are hard to compare because they mix different backbones, robot datasets, embodiments, preprocessing, and benchmark-specific tuning.
- That makes it unclear which parts of a system actually improve robot performance and which gains come from engineering around a benchmark.
- This matters because the field wants general robot policies, but fragmented evaluation hides whether methods transfer across tasks and embodiments.
Approach
- The authors build StarVLA-α, a simple baseline: a pretrained Qwen3-VL backbone plus a small MLP action head that predicts continuous action chunks from an action token.
- They keep the data pipeline minimal: raw RGB images, language instruction, training-split action normalization, and official benchmark evaluation without benchmark-specific tuning.
- They test the same setup across LIBERO, SimplerEnv, RoboTwin 2.0, and RoboCasa-GR1, and also train one joint generalist model with zero-padded actions up to dimension 32.
- Under matched backbone/data/training settings, they compare action-head choices: discrete FAST-style tokens, direct continuous regression, diffusion/flow-matching, and a dual-system GR00T-style design.
- They also ablate robot-data pretraining and common data engineering choices such as proprioception, history frames, delta actions, and relative actions.
Results
- Main benchmark results (specialist StarVLA-α): LIBERO 99.0 / 99.8 / 98.5 / 94.1, average 98.8 across Spatial/Object/Goal/Long; SimplerEnv 64.6 on WidowX, 70.2 on Google VA, 76.0 on Google VM; RoboTwin 2.0 50.3 clean, 88.2 clean*; RoboCasa-GR1 53.8. In the table, this beats OpenVLA-OFT average LIBERO 97.9, π0.5 RoboTwin clean 60.2 and clean* 82.7, and GR00T-N1.6 RoboCasa 47.6.
- The abstract claims the single generalist model beats π0.5 by 20% on the public real-world RoboChallenge benchmark, but the excerpt does not provide the raw benchmark numbers for that comparison.
- Action-head ablation: the simple MLP head is competitive with more complex continuous heads. On RoboCasa-GR1, MLP reaches 53.8, versus 52.8 for GR00T-style and 48.9 for diffusion-style π; discrete FAST is weaker at 45.0. On SimplerEnv Google VM, MLP gets 76.0 versus 60.1 for FAST.
- Robot pretraining ablation: extra action-data pretraining does not help consistently. Baseline StarVLA-α reaches RoboTwin clean 50.3 and RoboCasa 53.8; +OXE drops to 30.2 and 27.8; +InternData-A1 improves RoboTwin clean to 63.6 but lowers RoboCasa to 35.4; +RoboTwin-Rand raises RoboTwin clean to 79.7 but lowers RoboCasa to 33.3.
- Data engineering ablation: some additions help in low-data settings, but gains shrink with more task data. On RoboCasa 24×10, baseline is 9.8, delta action reaches 15.8, relative action 13.6, proprioception 12.5. With 24×1000, baseline is 53.8 and variants are close: proprioception 54.2, delta 54.8, relative 55.5.
- The paper’s claimed breakthrough is not a new architecture. It is the result that a strong pretrained VLM plus a small continuous action head can match or beat recent VLA systems across multiple benchmarks under one controlled recipe.