arXiv 2603.12510 ·
原版 PDF
·
项目页
·
代码Srikanth, Liang, Hsu, Bhatt, Zhao, Chen, Tjanaka, Hwang, Saran,
Seita, Tabrez, Nikolaidis · USC / Sony AI / Cornell
Q-DIG:让机器人听懂人话之前,先搜出它听不懂的人话
质量多样性指令生成 vs 视觉-语言-动作模型(arXiv 2603.12510)
阅读基线 arXiv 2603.12510v3;本页检索与版本注记更新于 2026-10-23。
先讲一个故事:会说人话的机器人,听不懂人话
一台真机机械臂站在桌前,任务是"push Coke can"——把可乐罐往前推。你说
"push Coke can",它完成得干净利落。你说 "meticulously exert force upon
the aluminum beverage
container"——同一个任务、同一罐可乐——它的夹爪在空中比划两下,什么都没碰到。九次里九次成功,十次里十次失败,差别只在措辞。
图 5(heatmap_no_colorbar2):一个真实 archive 热图——LIBERO-Goal
"put the bowl on top of the cabinet" 任务、OpenVLA-OFT。8 格 = 8
种攻击风格(z0–z7),格内数值 =
该风格精英指令的失败方差(0–1)。Q-DIG
在多种风格上都找到了高失败方差指令。
生成难例不等于判定是否安全。Metamorphic Testing of VLA Robots
通过输入变换间的执行关系缓解 oracle 问题;与 Q-DIG
的对抗样本搜索互补:前者提供关系型测试判据,后者定位模型对指令的脆弱区域。不能据此声称两者已组成经验证的联合系统。
Other Discussion Roles / 其他讨论角色
以下七个角色使用 MIT MAS.S60 课程原始
prompt,从不同责任位置审视同一篇论文。
Scientific Peer Reviewer
The paper has not been published yet and is currently submitted to a
top conference where you’ve been assigned as a peer reviewer.
Complete a full review of the paper answering all prompts of the
official review form of the top venue in this research area (e.g.,
NeurIPS). This includes recommending whether to accept or reject the
paper.
This paper was found buried under ground in the desert. You’re an
archeologist who must determine where this paper sits in the context
of previous and subsequent work. Find and report on one older paper
cited within the current paper that substantially influenced the
current paper and one newer paper that cites this current paper.
You’re a researcher who is working on a new project in this area.
Propose an imaginary follow-up project not just based on the current
but only possible due to the existence and success of the current
paper.
You work at a company or organization developing an application or
product of your choice (that has not already been suggested in a
prior session). Bring a convincing pitch for why you should be paid
to implement the method in the paper, and discuss at least one
positive and negative impact of this application.
You’re a hacker who needs a demo of this paper ASAP. Implement a
small part or simplified version of the paper on a small dataset or
toy problem. Prepare to share the core code of the algorithm to the
class and demo your implementation. Do not simply download and run
an existing implementation – though you are welcome to use (and give
credit to) an existing implementation for “backbone” code.
import random
random.seed(7)
TPL = "push the coke can"
STYLE = {
"z0": lambda s: "approach the can . apply force . stop .", # 分步指令
"z1": lambda s: "displace the aluminum beverage vessel", # 罕见词汇
"z2": lambda s: "hey buddy , could you push the coke can for me ?", # 人称呼唤
"z3": lambda s: s + " a bit", # 模糊副词
"z4": lambda s: s + " near the black marks", # 冗余细节
"z5": lambda s: "meticulously exert force upon the beverage container", # 冗长改写
"z6": lambda s: "yo , shove that can", # 口语俚语
"z7": lambda s: "move the can to the right of the fork", # 跨模态引用
}
TPL_TOK = set(TPL.split())
def p_fail(s): # toy VLA:只见过模板 → token 重叠决定成功率
tok = set(s.split())
overlap = len(tok & TPL_TOK) / max(len(tok), 1)
return max(0.02, min(0.98, 1 - overlap * 1.05 + random.uniform(-0.12, 0.12)))
def jac(a, b):
A, B = set(a.split()), set(b.split()); return 1 - len(A & B) / len(A | B)
def qdig(iters=600): # 条件化:每轮指定风格 z,格内留边界指令
arch = {}
for _ in range(iters):
z = random.choice(list(STYLE)); c = STYLE[z](TPL)
if 0.05 < 1 - p_fail(c) < 0.95: arch.setdefault(z, []).append(c)
return arch
def ert(iters=600): # 失败驱动爬山:从最失败候选出发随机再变异
pool, cur = [], TPL
for _ in range(iters):
c = STYLE[random.choice(list(STYLE))](cur) # 变异当前候选
if p_fail(c) >= p_fail(cur): cur = c # 只接受更失败的
if 0.05 < 1 - p_fail(c) < 0.95: pool.append(c)
return pool
qd, ert_pool = qdig(), ert()
qd_inst = [i for v in qd.values() for i in v]
ert_styles = {z for z in STYLE for i in [ert_pool] if any(s == STYLE[z](TPL) for s in i)}
pd_qd = [jac(a, b) for i, a in enumerate(qd_inst) for b in qd_inst[i+1:]]
pd_ert = [jac(a, b) for i, a in enumerate(ert_pool) for b in ert_pool[i+1:]]
print(f"Q-DIG:风格覆盖 {len(qd)}/8,成对 token 距离 {sum(pd_qd)/len(pd_qd):.3f}")
print(f"ERT :风格覆盖 {len(ert_styles)}/8,成对 token 距离 {sum(pd_ert)/len(pd_ert):.3f}")
You are a detective who needs to run a background check on one of
the paper’s authors. Where have they worked? What did they study?
What previous projects might have led to working on this one? What
motivated them to work on this project? Feel free to contact the
authors, but remember to be courteous, polite, and on-topic.
作者与机构(论文 v3 标注)。 11 位作者,USC 占 8
席:Siddharth Srikanth(一作/通讯,ssrikant@usc.edu)、Freddie
Liang、Ya-Chuan Hsu、Varun Bhatt、Shihan Zhao、Henry Chen、Bryon
Tjanaka、Minjune Hwang;Akanksha Saran(Sony
AI——注意七个月后的续作 Q-DEG 里她已列 Adobe
Research,以各论文署名时点为准);Daniel Seita、Aaquib
Tabrez(Cornell)与 Stefanos Nikolaidis(USC)三位标注 equal
advisement。Srikanth 的前作"算法化提示生成"(Algorithmic Prompt
Generation for Diverse Human-like Teaming and Communication with
LLMs,Srikanth et al. 2025;论文 related work 里称 PLAN-QD)做 LLM
智能体的 QD 提示搜索——一作的博士脉络从 LLM agent 延伸到 VLA。
调查者注。
我未联系任何作者;以上全部来自论文署名、致谢与公开学术页面。一点值得记录的谱系事实:Saran
在两篇之间的机构变动(Sony AI → Adobe)意味着这条"QD
红队"谱系的工业触点在扩大;而 DARPA EMHAT(Enhanced Measurement
for Hybrid Autonomy
Teaming)的资助说明语言红队在资助方眼里属于人机组队安全议题——这与方法本身的防御性定位一致。
Social Impact Assessor
Identify how this paper self-assesses its (likely positive) impact
on the world. Have any additional positive social impacts left out?
What are possible negative social impacts that were overlooked or
omitted?
误用面(有限)。
对已部署的语言控制机器人说对抗指令,后果是拒绝服务(任务失败),不是错误执行——Q-DIG
找的是让机器人失败的说法,不是让它做错事的说法。这限制了直接危害:最坏情形大致是"用一句话让服务机器人罢工"。真正需要警惕的反面情形论文没测:语言扰动是否也可能诱发不安全的成功(模糊副词"push
it a
bit"导致的过度位移、跨模态引用导致的抓错物体)——失败与错误执行是两种不同的风险面,本文只覆盖前者。
直接前作 ERT(攻击风格 taxonomy z0–z5 来源):Karnik et al.
Embodied Red Teaming for Auditing Robotic Foundation Models, 2024
QD × LLM 提示轴:Samvelyan et al.
Rainbow Teaming: Open-Ended Generation of Diverse Adversarial
Prompts, NeurIPS 2024;Srikanth et al.
Algorithmic Prompt Generation for Diverse Human-like Teaming and
Communication with Large Language Models, 2025(论文中称 PLAN-QD)
QD 基础:Lehman & Stanley,
Abandoning Objectives: Evolution Through the Search for Novelty
Alone, 2011;Mouret & Clune,
Illuminating Search Spaces by Mapping Elites, 2015;Pugh,
Pace & Stanley,
Confronting the Challenge of Quality Diversity,
2015;Fontaine & Nikolaidis,
A Quality Diversity Approach to Automatically Generating
Human-Robot Interaction Scenarios in Shared Autonomy, RSS 2021;Bhatt et al.,
Surrogate Assisted Generation of Human-Robot Interaction
Scenarios, 2023
VLA 脆弱性与模型:pi0-experiment-wild(Black et al.,
指令敏感性系统测量);OpenVLA-OFT(Kim et al.,
2025);π₀/π₀.₅(Black et al.);GR00T
N1.5/N1.6(NVIDIA);LIBERO(Liu et al., 2023);SimplerEnv(Li et
al., 2024)