PRTS: A Primitive Reasoning and Tasking System via Contrastive Representations

Researchers introduce PRTS, a Vision-Language-Action foundation model that reformulates robotic learning through goal-conditioned reinforcement learning rather than traditional behavior cloning. The system learns to assess goal reachability by embedding state-action pairs and language instructions in a unified space, achieving state-of-the-art performance on multiple robotic benchmarks and real-world tasks.

PRTS represents a meaningful shift in how foundation models approach robotic control by incorporating temporal reasoning into the pretraining process. Rather than treating robot learning as simple imitation, the system frames it as a goal-reaching problem where the model must understand not just what action to perform, but whether that action brings the robot closer to completing a language-specified objective. This distinction becomes critical in long-horizon tasks where intermediate steps must logically progress toward an end goal.

The technical innovation leverages contrastive reinforcement learning to create dense supervision signals directly from offline trajectory data without requiring explicit reward annotations. By computing the inner product between state-action and goal embeddings as a proxy for goal occupancy probability, PRTS bridges the gap between high-level semantic reasoning and low-level physical feasibility. The approach integrates seamlessly into existing vision-language model architectures through role-aware causal masking, avoiding significant computational overhead compared to standard behavior cloning.

The empirical validation across LIBERO, SimplerEnv, and 14 real-world manipulation tasks demonstrates tangible improvements, particularly in zero-shot generalization and contact-rich scenarios where naive imitation learning typically struggles. These gains suggest that explicit goal-reachability awareness helps models understand task constraints and physical limitations beyond what semantic matching alone provides.

For the robotics and embodied AI community, PRTS establishes a blueprint for incorporating reinforcement learning principles into foundation model pretraining at scale. The work validates that injecting temporal task progress awareness into VLA models produces more capable and generalizable policies, potentially influencing future robotics research methodologies.

• →PRTS reformulates robotic pretraining through goal-conditioned reinforcement learning rather than behavior cloning, incorporating temporal task progress awareness.
• →The system learns goal reachability by matching state-action and language embeddings without explicit reward annotations, extracting supervision from offline trajectory data.
• →State-of-the-art results on LIBERO, SimplerEnv, and real-world benchmarks show substantial improvements in long-horizon, contact-rich, and zero-shot settings.
• →Integration into VLM architectures requires minimal computational overhead through role-aware causal masking, making adoption practical for existing systems.
• →The work demonstrates that understanding physical feasibility and goal reachability significantly improves both execution success and planning capabilities of robotic policies.