Vision-Language-Action (VLA) models translate natural-language commands into robot action sequences. However, leading systems on the LIBERO-Plus robustness benchmark utilize backbones with three to seven billion parameters, which can exceed the memory budget of embedded robots. We present CoTinyVLA, a 0.9B-parameter action model based on Qwen3.5-0.8B, achieving robustness through structured supervision instead of model enlargement.
CoTinyVLA comprises three key components targeting different aspects of the problem:
- Dual-view temporal input: 16 historical frames per step with textual camera and time markers;
- Hierarchical chain-of-thought (CoT) distillation: distillation from a 35B teacher model into an episode-level Plan and chunk-level Think over task phases, gripper states, and next subactions;
- Paraphrase augmentation: expanding 40 base commands into 800 variants.
On LIBERO-Plus, covering 10,030 perturbed tasks across seven perturbation dimensions, CoTinyVLA achieved 90.8% on Spatial, 87.3% on Object, 86.6% on Goal, and 80.7% on Long tasks, outperforming the strongest 7B baseline by 4.7, 2.8, 15.9, and 3.0 points respectively, with all margins excluding zero. Gains are concentrated on the hardest axes of the benchmark: none of the eleven published baselines exceed 53.2% on Robot Initial States, while CoTinyVLA reached 73.6% on Goal against 39.9% for the strongest baseline. Ablations show that the three components are separable by perturbation axis, and at a matched image budget, how frames are divided between the two cameras and across time accounts for an 8.6-point contribution alone. Closed-loop inference peaks at 2.25 GiB of allocated GPU memory, and paired interventions indicate that the episode Plan is load-bearing: replacing it with an empty or contradictory span costs 40 to 45 points of success. Thus, structured supervision allows a 0.9B backbone to exceed all baselines.
Code link: CoTinyVLA GitHub
Blogger's Review: CoTinyVLA demonstrates the potential to significantly enhance the performance of small models through structured supervision, especially in resource-constrained environments. Its innovative chain-of-thought distillation method offers new insights for future robotic learning, warranting further exploration and application.