TeleHuman/PRTS

What it solves

PRTS (Primitive Reasoning and Tasking System) addresses the limitation of standard Vision-Language-Action (VLA) models that rely solely on behavior cloning. While traditional models learn what to do, they often lack an internal understanding of how close the current state is to achieving the goal. PRTS introduces a way for models to internalize a sense of goal-reachability, allowing them to follow novel instructions and recover from interventions more effectively than models trained only on actions.

How it works

PRTS scales reward-free contrastive reinforcement learning (RL) into the pre-training phase of a VLA model. It uses a Qwen3-VL backbone and co-trains a contrastive value head alongside the action loss. Instead of requiring curated reward labels or success markers, it derives supervision purely from the temporal structure of offline demonstration trajectories. This allows the model to learn a language-grounded value function where the inner product of state-action and goal embeddings tracks the goal-occupancy, rising as the policy approaches the target.

Who it’s for

Robotics researchers and developers working with Vision-Language-Action models who need better generalization to novel instructions, long-horizon task execution, and robustness in real-world deployments across different robot platforms (e.g., dual-arm or single-arm systems).

Highlights

• Reward-Label-Free: Learns goal-reachability without needing per-episode success labels or curated reward datasets.
• Goal-Reachability Awareness: Integrates a contrastive value head directly into the backbone, enabling the model to understand progress toward a goal.
• High Efficiency: Uses a custom CuTe-FlashAttention kernel to keep pre-training compute costs near those of vanilla behavior cloning.
• Strong Generalization: Significantly outperforms prior VLAs on novel-instruction following and out-of-distribution tasks in both simulation and real-world environments.