# Stacking 10x More Data Still Loses by 40% -- Psi-Zero Says Mixed Training for Humanoid Robots Is Simply Wrong

Policy Learning with Limited Robot Data

Full article: https://haiguangboy.com/posts/psi-zero

## Core Method

Ψ0 splits the training objective into three stages:
· Stage 1: Pretrain Qwen3-VL-2B-Instruct on ~829 hours of EgoDex human first-person video, predicting only single-step action tokens, focusing on learning visual representations and task semantics
· Stage 2: Freeze the VLM, and train a separate MM-DiT flow-matching action expert on ~3 million frames / 30 hours of real humanoid robot data
· Stage 3: For each downstream long-horizon task, use only 80 teleoperation demonstrations to fine-tune the action expert for 40,000 steps

It also replaces the unstable RTC at test time with training-time RTC: during training, randomly mask the first 1-6 action tokens, teaching the model to continue subsequent action blocks conditioned on prior clean actions. At inference, unexecuted actions can serve as clean tokens for continued conditioning, reducing jitter from "pause-think-execute".

## Key Results

The most striking result is data efficiency: Ψ0 uses only ~800 hours of human video and 30 hours of robot data, achieving an overall success rate over 40% higher than the second-place GR00T N1.6 across 8 real-world long-horizon precision manipulation tasks, while other baselines used over 10x more data.

Ablations also support the necessity of the three-stage decoupling. Directly freezing Qwen3-VL and fine-tuning the action head yields an overall success rate of only 0.2; adding EgoDex pretraining brings a clear improvement; further adding real robot post-training and training-time RTC continues to stack contributions.

This shows that the value of pretraining lies not in how accurately actions are predicted, but in first learning the visual-task alignment. Action dynamics are left for later robot data and the action expert to learn separately.

## Comparison with Related Approaches

- No structured associations yet.

## Limitations

This is not "robot data doesn't matter." Ψ0 still requires high-quality real robot data for post-training the action expert, and also needs a small number of demonstrations per downstream task for adaptation.

Nor does it prove that all robot tasks can be solved with limited data. The experiments focus on long-horizon precision manipulation and mobile manipulation for humanoid robots; hardware, teleoperation frameworks, VLM scale, and training compute all contribute to the results. The promise of open models and training pipelines also depends on future releases.

## Author's Assessment (Not from the paper; cross-paper synthesis)

What makes Ψ0 truly valuable is not building another humanoid robot foundation model, but clarifying why limited robot data can be effective: first use human video to learn visual representations and task semantics, then use robot data to learn action dynamics, rather than forcing the two distributions into a single end-to-end objective.

This will influence future VLA engineering trade-offs. Data scale certainly matters, but what matters more is the role of data within the training objective: human video is suited for shaping representations, robot data is suited for calibrating actions. By separating these responsibilities, limited real data can be amplified.