ActionEQA: Action Interface for Embodied Question Answering

Tianwei Bao^1*  ·  Qineng Wang^1*  ·  Kangrui Wang¹  ·  Mingkai Deng²  ·  Guangyi Liu⁵  ·  Jiayuan Mao³
Larry Birnbaum¹  ·  Zhiting Hu⁴  ·  Eric P. Xing^2,5  ·  Zhaoran Wang¹  ·  Manling Li¹

¹ Northwestern University     ² Carnegie Mellon University     ³ UPenn
⁴ UC San Diego     ⁵ MBZUAI

^{* Equal contribution}

ActionEQA is the first action-centric Embodied Question Answering (EQA) benchmark designed to systematically evaluate the ability of Vision-Language Models (VLMs) to bridge the semantic-to-physical gap: translating high-level semantic instructions into precise low-level physical robot actions.

Overview

A pivotal challenge for embodied agents is bridging the semantic-to-physical gap: translating abstract goals (the "what") into the precise motor commands required for physical interaction (the "how"). Existing benchmarks focus on high-level perception and planning, failing to capture the depth and nature of this divide.

ActionEQA addresses this with two core design principles:

1. Three-Tiered Action Hierarchy

Actions are decomposed into three levels of abstraction:

Level	Symbol	Description	Example
High	a_high	Natural language goal — the "what"	"Close the Microwave"
Mid	a_mid	Semantic motion description — the "semantic how"	"Move along positive X direction, rotate clockwise around Z-axis"
Low	a_low	Raw 7-DoF end-effector command — the "physical how"	[Δx, Δy, Δz, Δroll, Δpitch, Δyaw, Δgripper]

2. Bidirectional Reasoning Framework

Each action level is evaluated in two complementary directions:

Task	Direction	Given	Predict
State Prediction (SP)	Forward	Initial state s_t + action a_t	Resulting state s_{t+1}
Action Inference (AI)	Backward	Initial state s_t + final state s_{t+1}	Action a_t that caused the transition

These two principles combine to yield 6 evaluation categories: H-SP, H-AI, M-SP, M-AI, L-SP, L-AI.

---

Dataset Statistics

Statistic	DROID	BridgeData V2	RT-1	Total
Questions	2,953	4,732	1,110	8,795
Unique Images	8,026	14,146	4,144	26,213
Unique Episodes	367	1,707	555	2,629

Questions by Category

Category	Description	Count
H-SP	High-Level State Prediction	1,891
H-AI	High-Level Action Inference	1,891
M-SP	Mid-Level State Prediction	947
M-AI	Mid-Level Action Inference	1,379
L-SP	Low-Level State Prediction	947
L-AI	Low-Level Action Inference	1,740
Total		8,795

Note: RT-1 is used only for high-level tasks (H-SP and H-AI) because its mobile base introduces confounds for mid/low-level action analysis.

---

Dataset Configurations

Each configuration name follows the pattern: {source}_{level}_{task}.

Component	Values	Meaning
source	bridge, droid, rt1	Source robotics dataset
level	high, mid, low	Action hierarchy level
task	forward, inverse	State Prediction (forward) or Action Inference (inverse)

Available configs:

Config	Task	Source
bridge_high_forward	High-Level State Prediction	BridgeData V2
bridge_high_inverse	High-Level Action Inference	BridgeData V2
bridge_mid_forward	Mid-Level State Prediction	BridgeData V2
bridge_mid_inverse	Mid-Level Action Inference	BridgeData V2
bridge_low_forward	Low-Level State Prediction	BridgeData V2
bridge_low_inverse	Low-Level Action Inference	BridgeData V2
droid_high_forward	High-Level State Prediction	DROID
droid_high_inverse	High-Level Action Inference	DROID
droid_mid_forward	Mid-Level State Prediction	DROID
droid_mid_inverse	Mid-Level Action Inference	DROID
droid_low_forward	Low-Level State Prediction	DROID
droid_low_inverse	Low-Level Action Inference	DROID
rt1_high_forward	High-Level State Prediction	RT-1
rt1_high_inverse	High-Level Action Inference	RT-1

---

Data Schema

State Prediction (*_forward) — Forward Task

Given an initial state and an action description, select the correct resulting state from 4 image candidates.

Field	Type	Description
frame1	Image	Initial state s_t
action	string	Action description (NL goal / semantic motion / 7-DoF vector)
option_A	Image	Candidate resulting state A
option_B	Image	Candidate resulting state B
option_C	Image	Candidate resulting state C
option_D	Image	Candidate resulting state D
correct_ans	ClassLabel	Ground-truth answer: one of A, B, C, D

Action Inference (*_inverse) — Backward Task

Given before and after states, select the correct action from 4 candidates.

Field	Type	Description
frame1	Image	Before state s_t
frame2	Image	After state s_{t+1}
option_A	string	Candidate action A (NL description / motion label / 7-DoF vector)
option_B	string	Candidate action B
option_C	string	Candidate action C
option_D	string	Candidate action D
correct_ans	ClassLabel	Ground-truth answer: one of A, B, C, D

---

Usage

from datasets import load_dataset

# High-Level State Prediction — BridgeData V2
ds = load_dataset("TianweiBao/ActionEQA", "bridge_high_forward", split="train")

# High-Level Action Inference — DROID
ds = load_dataset("TianweiBao/ActionEQA", "droid_high_inverse", split="train")

# Mid-Level State Prediction — BridgeData V2
ds = load_dataset("TianweiBao/ActionEQA", "bridge_mid_forward", split="train")

# Low-Level Action Inference — DROID
ds = load_dataset("TianweiBao/ActionEQA", "droid_low_inverse", split="train")

# Inspect a sample
sample = ds[0]
print("Before state:", sample["frame1"])   # PIL Image
print("After state:",  sample["frame2"])   # PIL Image
print("Options:", sample["option_A"], sample["option_B"], sample["option_C"], sample["option_D"])
print("Correct answer:", sample["correct_ans"])

Load all configs at once

from datasets import load_dataset

configs = [
    "bridge_high_forward", "bridge_high_inverse",
    "bridge_mid_forward",  "bridge_mid_inverse",
    "bridge_low_forward",  "bridge_low_inverse",
    "droid_high_forward",  "droid_high_inverse",
    "droid_mid_forward",   "droid_mid_inverse",
    "droid_low_forward",   "droid_low_inverse",
    "rt1_high_forward",    "rt1_high_inverse",
]

datasets = {cfg: load_dataset("TianweiBao/ActionEQA", cfg, split="train") for cfg in configs}

---

Citation

If ActionEQA is useful for your research, please cite:

@article{bao2026actioneqa,
  title   = {ActionEQA: Action Interface for Embodied Question Answering},
  author  = {Bao, Tianwei and Wang, Qineng and Wang, Kangrui and Deng, Mingkai
             and Liu, Guangyi and Mao, Jiayuan and Birnbaum, Larry and Hu, Zhiting
             and Xing, Eric P. and Wang, Zhaoran and Li, Manling},
  journal = {Transactions on Machine Learning Research},
  year    = {2026},
  url     = {https://openreview.net/forum?id=HY2ruqdMt4}
}

---

License

ActionEQA is built on top of DROID, BridgeData V2, and RT-1. Please refer to the respective dataset licenses for terms of use.