Prompting Decision Transformers for Zero-Shot Reach-Avoid Policies.

Area of Science:

• Artificial Intelligence
• Robotics
• Reinforcement Learning

Background:

• Offline goal-conditioned reinforcement learning (RL) is effective for reach-avoid tasks.
• Current methods struggle with dynamic avoid-region specification and require well-designed reward functions.
• This limits scalability in complex environments.

Purpose of the Study:

• Introduce RADT, a decision transformer for offline, reward-free, goal-conditioned, and avoid-region-conditioned RL.
• Enable flexible, dynamic specification of multiple avoid regions at evaluation time.
• Learn effective reach-avoid policies from suboptimal data without explicit rewards.

Main Methods:

• RADT encodes goals and avoid regions as prompt tokens within a decision transformer architecture.
• Utilizes a novel combination of goal and avoid-region hindsight relabeling.
• Trained on suboptimal offline trajectories from a random policy.

Main Results:

• RADT demonstrates zero-shot generalization to out-of-distribution avoid region sizes and counts.
• Outperforms existing offline goal-conditioned RL models across 11 diverse tasks.
• Achieves significant improvement in normalized cost (35.7%) while maintaining high success rates.

Conclusions:

• RADT offers a flexible and scalable approach to reach-avoid tasks in complex, poorly structured environments.
• Its ability to handle dynamic avoid regions and reward-free learning is a significant advancement.
• Successfully applied to biological cell reprogramming, reducing undesirable state visits.