Decision Transformer Pocket

Decision Transformer Pocket trains an offline return-conditioned policy on 4,000 mixed-quality corridor trajectories. A nearby claim pays 0.4; the larger reward requires first moving away from it, retrieving a key, crossing a door, and claiming treasure.

The behavior-cloning control sees identical states and actions but no requested return. Evaluation asks each policy to realize both low- and high-return targets from three starting positions.

Verified local result

The 18,371-parameter Decision Transformer selected the nearby reward in 100% of 300 target-0.4 episodes and the key-door treasure in 100% of 300 target-1.0 episodes. The 595-parameter behavior-cloning policy could not switch intention, reaching only 33.3% and 66.7% desired-terminal rates.

uv run python projects/decision-transformer-pocket/train.py
uv run pytest tests/test_decision_transformer_pocket.py

Hosted showcase

This free static Space preserves the complete original Gradio source, trained artifacts, evaluation files, and local launch requirements. Hugging Face now requires PRO for CPU-backed Gradio hosting, so the public landing page is static while the checked-in app.py remains the authoritative runnable demo.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support