🤖🧠 Embodied Agent Interface (EAI)

Benchmarking LLMs for Embodied Reasoning and Decision Making with Safety Constraints

🔗 Original repository:
https://github.com/embodied-agent-interface/embodied-agent-interface

---

📌 Overview

This benchmark provides a comprehensive assessment of LLM performance across different subtasks, identifying their strengths and weaknesses in embodied decision-making contexts with safety constraints.

It includes:

1. ✅ Tasks that can be achieved safely under adversarial and multi-constraint settings
2. ⚠️ Adversarial instructions that the agent must avoid

---

🛠️ Running the Experiments

We consider two planning strategies:

• 🚀 One-go Planning → ideal for direct planning
• 🔁 Stepwise Planning → ideal for iterative planning

---

🚀🧠 One-go Planning

In the one-go planning approach, the agent 🤖 generates a plan 📋 (a sequence of actions) in a single attempt for a given instruction or task.
▶️ The plan is then executed by the simulator 🎮.
📊 Upon successful execution, the environment graph 🗺️ 𝒢* (representing the updated state of the environment) is evaluated.

💡 Thanks to its straightforward design, this planning strategy is well-suited for direct planning scenarios.

▶️ Example Command

python src/behavior_eval/evaluation/action_sequencing/scripts/evaluate_results.py \
    --mode onego \
    --llm_name LGAI-EXAONE/EXAONE-3.5-32B-Instruct \
    --strategy direct

---

🔁 Stepwise Planning

In stepwise planning, the agent 🤖 interacts with the environment 🌍 for n steps and m trials to finish the task.
At each step, the agent selects and executes an action $a_{ij}$, where $i \in m$ and $j \in n$, which is processed by the simulator 🎮 and returns an observation $o_{ij}$, along with the updated environment state $\mathcal{G}_{ij}$.

This interaction continues for a (pre-defined) number of steps, constituting a trajectory:

$$ \tau_i = { a_{11}, o_{11}, a_{12}, o_{12}, \ldots } $$

At the end of each trial, a critic 🧠 evaluates the (so-far generated) plan $\mathcal{P}_i$ and provides feedback $f_i$.
This feedback guides the agent in refining its strategy or decision-making process for future trials.

The process continues until the agent generates the “Done” action or exhausts all trials.
If the plan is executable, the updated environment state $\mathcal{G}^*$ is passed on for evaluation.

✨ This strategy suits iterative planning, i.e., where interactions occur between the agent and the simulator/environment.

⚙️ Key Arguments

`--llm_name`: `Qwen/Qwen3-32B`  `gpt-4.1-mini` etc 
`--strategy`: `direct`  `react` `rej`
`--rm_safety_instruction`: if set, then instruction for safety enhancement in the prompt will be removed, for ablation study.
`--reflex`: if set, `reflexion` or `AI-critic from gpt-4.1` will be applied.
`--reflex_from_llm_as_judge`: if set, use `AI-critic from gpt-4.1`, otherwise, use `reflexion` 
`--trial`: number of rounds for `reflexion` or `AI-critic`

🧪 Example Experiments

🔁 Stepwise + ReAct

python src/behavior_eval/evaluation/action_sequencing/scripts/evaluate_results.py \
--mode stepwise --llm_name ${llm}  --strategy react 

🧠 Stepwise + ReAct + Reflexion / AI-Critic

# reflexion for additional N rounds (N is 1,2,3,...)
python src/behavior_eval/evaluation/action_sequencing/scripts/evaluate_results.py \
--mode stepwise --llm_name ${llm}  --strategy react \
--reflex_from_llm_as_judge \
--trial 1 

🧹 Ablation: Remove Safety Instruction

python src/behavior_eval/evaluation/action_sequencing/scripts/evaluate_results.py \
--mode onego --strategy direct --llm_name ${llm}    \
--rm_safety_instruction

python src/behavior_eval/evaluation/action_sequencing/scripts/evaluate_results.py \
--mode stepwise --strategy direct --llm_name ${llm}    \
--rm_safety_instruction 

⚠️ Risk-Specific Evaluation (e.g., Electrical Hazard)

python src/behavior_eval/evaluation/action_sequencing/scripts/evaluate_results.py \
--mode stepwise --strategy react --llm_name ${llm}   \
--reflex_from_llm_as_judge \
--trial 1 \
--risk_category 'electrical hazard' 

Citation

If you find this work helpful, please consider citing it:

@inproceedings{sadhu2025vestabench,
  title={VESTABENCH: An Embodied Benchmark for Safe Long-Horizon Planning Under Multi-Constraint and Adversarial Settings},
  author={Sadhu, Tanmana and Chen, Yanan and Pesaranghader, Ali},
  booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track},
  year={2025}
}

License

This project is released under the MIT License.