Summary

PRTS addresses the challenge of scaling reward-free contrastive reinforcement learning (RL) into Vision-Language-Action (VLA) model pre-training. It equips a Qwen3-VL backbone with a quantitative, language-grounded sense of goal-reachability using only offline trajectory data, enabling robust embodied AI agents with near-Behavioral Cloning compute costs.

How It Works

PRTS reframes VLA pre-training as a goal-conditioned RL problem, supervising a language-conditioned contrastive value alongside the action loss. This approach, derived solely from offline trajectory structure, avoids curated reward datasets or separate value networks. The model learns a sharp geometric representation where state-action and goal embeddings approximate log-discounted goal-occupancy, indicating progress towards language goals.

Quick Start & Requirements

• Environment: CUDA 12.6, PyTorch 2.11, transformers 4.57.3.
• Installation: Clone repo, create conda env (Python 3.11), install LeRobot (pip install lerobot==0.3.3), dependencies (pip install -r requirements.txt), FlashAttention (pip install flash-attn==2.8.3 --no-build-isolation), and PRTS (pip install -e .).
• Pre-trained Models: Download TeleEmbodied/PRTS-4B and Qwen/Qwen3-VL-4B-Instruct via huggingface-cli.
• Links: Project Page: https://rhodes-team-prts.github.io/, HuggingFace: https://huggingface.co/TeleEmbodied/PRTS-4B, arXiv: https://arxiv.org/abs/2604.27472.

Highlighted Details

• PRTS-Droid ranks 3rd on the MolmoSpaces Leaderboard (42.4% SR), outperforming NVIDIA's DreamZero.
• Achieves state-of-the-art on LIBERO, LIBERO-Plus, LIBERO-Pro, and SimplerEnv benchmarks with significantly less post-training compute.
• Demonstrates strong out-of-distribution generalization, particularly in novel instruction following and long-horizon tasks, outperforming π 0.5 by large margins.
• Achieves ≥ 90% success on real-world dual-arm (RealMan) and single-arm (Flexiv) robotic tasks.
• Pre-training scales efficiently (≥ 85% linear scaling up to 64 GPUs) with a custom FlashAttention kernel.

Maintenance & Community

• Contact: Yang Zhang (breezeyoung9470@gmail.com) for collaborations or inquiries.
• Community: WeChat discussion group available.
• Development: Ongoing progressive open-sourcing of the PRTS stack.

Licensing & Compatibility

• License: CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0).
• Restrictions: Free for academic and non-commercial use; commercial use is prohibited.

Limitations & Caveats

• The CC BY-NC 4.0 license strictly prohibits commercial use.
• Some advanced scripts, like CRL value visualization, are still upcoming.
• Requires specific, recent versions of CUDA (12.6) and PyTorch (2.11).