機器人／邊緣 AI
CoTinyVLA Outperforms 7B Robotics Baselines With Just 0.9B Parameters, but Testing Remains Limited to Simulation
CoTinyVLA distills hierarchical reasoning from a 35B teacher into a 0.9B vision-language-action model, outperforming the strongest 7B baseline across all four LIBERO-Plus test suites. Its code, training pipeline, and weights are public, but there are still no results from real robots, different hardware, or independent reproductions.
Robotics VLA models typically seek better generalization by scaling up their language backbones. CoTinyVLA instead attempts to compress the necessary reasoning structure into 0.9B parameters. The system uses Qwen3.5-0.8B as its backbone. At each control step, it processes a 16-frame history from third-person and wrist-mounted cameras, augmented with camera and temporal markers to prevent the model from relying solely on the latest frame. During training, a 35B teacher first generates two levels of textual supervision: a task-level `Plan` and a `Think` trace describing the current stage, gripper state, and next sub-action. The original 40 instructions are also expanded into 800 phrasings.
On LIBERO-Plus, which contains 10,030 perturbed tasks, CoTinyVLA achieves 90.8%, 87.3%, 86.6%, and 80.7% on the Spatial, Object, Goal, and Long suites, respectively. These scores exceed the strongest existing 7B result for each suite by 2.8 to 15.9 percentage points. The largest margin appears under changes to the robot’s initial state, suggesting that temporal history and structured supervision may improve state recognition in closed-loop control more effectively than simply adding parameters.
The authors report a peak configuration memory footprint of 2.25 GiB during closed-loop inference. On an L40S with three concurrent rollouts per GPU, steady-state latency is 1.37 seconds per eight actions. Caching the episode-level `Plan` reduces steady-state latency by 48.5%. Removing or replacing the `Plan` at test time lowers the success rate by 40 to 45 percentage points, indicating that the textual reasoning is not merely decorative output.
However, the results come from the LIBERO simulator, and each LIBERO-Plus perturbation task is run only once. They therefore cannot establish reliability on physical robot arms exposed to sensor noise, control latency, and hardware wear. Although the repository includes data conversion, training, and evaluation pipelines, community adoption and independent validation remain nearly nonexistent. The next questions are how well the model transfers to real robots, how sensitive distillation is to the choice of teacher model, and what control frequency it can achieve in practice on embedded GPUs.