Re^2 Agent: Reflection and Re-execution Agent for Embodied Decision Making

This repository Re2_Agent contains example code used by nju-lamda12 for the Embodied Agent Interface (EAI) Challenge.

We focus on:

• Optimized prompt templates for EAI-style tasks.
• Scripts to batch-process prompts and call LLM OpenAI.
• An agent-reflection module that reads execution logs and asks an LLM to fix task.

---

Structure

• stage1_prompt/

  • Prompt templates (e.g., for behavior and VirtualHome tasks).
  • process_prompt.py: batch prompt processing + LLM calling.

• stage2_agent/

  • data/virtualhome_report/: example VirtualHome evaluation logs.
  • Re-excution_agent.py: example re-execution / reflection agent that:
    • reads prompts and old outputs,
    • parses execution logs,
    • lets the LLM output a corrected JSON action sequence.

---

How to Use (Typical Example)

From the project root:

cd /home/eai-eval/Re2_Agent

# 1) Generate initial action sequences (example)
python stage1_prompt/process_prompt.py \
  --input_file ./prompt/behavior_action_sequencing_prompts.json \
  --output_file ./output/behavior_action_sequencing_outputs.json \
  --model YOUR_API_DEPLOYMENT_NAME

# 2) After running evaluation scripts and getting logs,
#    run the reflection agent for VirtualHome:
python stage2_agent/re-excution_agent.py \
  --old_output_file virtualhome/virtualhome_action_sequencing_outputs.json \
  --eval_log_file data/virtualhome_report/logs/action_sequencing_eval_20251123_232824.log \
  --output_file virtualhome_v2/virtualhome_action_sequencing_outputs_FIXED.json

You may need to adjust paths, model names, and environment variables (e.g., for Azure OpenAI) according to your local setup.

---

Acknowledgements

We thank the Embodied Agent Interface (EAI) organizers and platform for providing the benchmark, tools, and evaluation environment that made this project possible.