Auriga

A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation

面向零样本具身操作的任务无关智能体框架

A persistent model session understands the task, calls tools and controls the robot, revising its plan from real execution feedback. Every physical action passes through one execution gateway, and every run leaves a replayable evidence trail.

一个持续运行的模型会话负责理解任务、调用工具并控制机器人，并根据真实的执行反馈修订计划。所有物理动作都经过同一个执行网关，每次运行都会留下可回放的证据记录。

StellarEdge Team

Code代码 Tech Report技术报告 Demos演示
3476 / 6300episodes solved集成功 0.640dimension-macro score维度等权均分 42task families类任务
background: Auriga on arrange_auriga (custom demo task: spell AURIGA), head camera, HD replay背景：Auriga 执行 arrange_auriga（自建演示任务：拼出 AURIGA），头部相机，高清回放
Results结果

RoboDojo resultsRoboDojo 实验结果

Auriga with GPT-6 astra (high reasoning), evaluated on 42 tasks with 50 layouts per task across three seeds under the official RoboDojo simulation settings.

Auriga + GPT-6 astra（high 推理），按照 RoboDojo 官方仿真评测设置，在 42 个任务、每任务 50 个 layout 上运行三个 seed。

57.7%dimension-macro success rate维度等权成功率
0.640dimension-macro mean score / 1维度等权均分 / 1
Harness · model框架 · 模型
Auriga · GPT-6 astra · high reasoningAuriga · GPT-6 astra · high 推理档位
Protocol评测设置
42 tasks × 50 layouts × 3 seeds (0, 1, 2); 6,300 episodes.42 个任务 × 50 个 layout × 3 个 seed（0、1、2），共 6300 局。
Environment环境
Official RoboDojo simulation settings: Isaac Sim, ARX X5 dual arm, native step limits and native evaluator.与 RoboDojo 官方仿真评测设置一致：Isaac Sim、ARX X5 双臂、原生步数上限和原生评估器。
Aggregation汇总方式
Average tasks equally within each dimension, then average the five dimensions and three seeds equally. Scores use the 0–1 scale.维度内任务等权，五个维度等权，三个 seed 等权；分数采用 0–1 标度。
Missing episodes缺失局处理
Unscored or invalid episodes count as failures with score 0. All 6,300 planned episodes remain in the denominator.未评分或无效局按失败、0 分计入，分母固定为计划的 6300 局。

Pooled episode success: 3,476 / 6,300 (55.17%). This weights every episode equally and differs from the dimension-macro headline.

按局数合计：3476 / 6300（55.17%）。此口径对每局等权，与顶部的维度等权指标不同。

Results by capability dimension按能力维度的结果
Dimension能力维度Tasks任务数Successful / episodes成功 / 局数Success rate成功率Mean score / 1均分 / 1
Memory记忆6835/90092.8%0.932
Precision精细操作8462/120038.5%0.463
Long-Horizon长程8542/120045.2%0.620
Open开放指令8757/120063.1%0.654
Generalization泛化12880/180048.9%0.528
Results by seed分 seed 结果
SeedSuccessful / episodes成功 / 局数Dimension-macro success维度等权成功率Dimension-macro score维度等权均分
01170/210058.1%0.642
11168/210058.1%0.643
21138/210056.9%0.634
Per-task results · 3 seeds × 50 layouts逐任务结果 · 3 个 seed × 50 个 layout
Task任务Successful / 150成功 / 150Success rate成功率Mean score / 1均分 / 1
Memory记忆
match_and_pick_from_conveyor145/15096.7%0.967
swap_blocks145/15096.7%0.967
imitate_sorting_sequence141/15094.0%0.953
press_by_number142/15094.7%0.947
swap_T137/15091.3%0.913
cover_blocks125/15083.3%0.846
Open开放指令
stack_blocks_by_language144/15096.0%0.968
general_pickup140/15093.3%0.933
classify_objects_by_language133/15088.7%0.921
align_blocks136/15090.7%0.907
solve_equation131/15087.3%0.873
pick_from_conveyor_by_image67/15044.7%0.447
store_tools_in_toolbox0/1500.0%0.133
pour_by_language6/1504.0%0.049
Precision精细操作
fasten_screws104/15069.3%0.807
play_Xylophone116/15077.3%0.773
insert_tubes69/15046.0%0.613
deposit_coin48/15032.0%0.425
pour_balls_into_vase57/15038.0%0.380
build_tower37/15024.7%0.352
plug_in_charger31/15020.7%0.207
insert_key0/1500.0%0.148
Task任务Successful / 150成功 / 150Success rate成功率Mean score / 1均分 / 1
Generalization泛化
stack_blocks147/15098.0%0.983
stack_bowls137/15091.3%0.925
arrange_largest_number132/15088.0%0.909
push_T129/15086.0%0.860
fold_clothes104/15069.3%0.745
sort_nesting_dolls_by_size92/15061.3%0.613
sweep_blocks61/15040.7%0.407
make_toast39/15026.0%0.287
pack_objects_into_box22/15014.7%0.273
hang_mugs6/1504.0%0.195
pour_liquid_into_cup11/1507.3%0.073
store_laptop_and_headphones0/1500.0%0.071
Long-Horizon长程
classify_objects145/15096.7%0.977
play_tic_tac_toe125/15083.3%0.944
put_bottles_into_dustbin124/15082.7%0.893
organize_table29/15019.3%0.690
play_stacking_toy54/15036.0%0.647
fill_pen_holder23/15015.3%0.430
make_kong41/15027.3%0.273
fill_egg_holder1/1500.7%0.110

Task-by-task outcomes逐任务执行结果

Each square represents one episode. Within each task and seed, squares are ordered by outcome: green successes, then gray failures. This visualizes the aggregate counts; positions do not identify individual layouts.

每格代表一局。各任务、各 seed 内按结果排列：绿色成功在前，浅灰失败在后。这张图展示汇总数量，方格位置不对应具体 layout 编号。

Open full size ↗查看原图 ↗Download SVG ↓下载 SVG ↓
Auriga RoboDojo: 42 tasks, 3 seeds, 6300 outcome-sorted episode squares; 3476 successes. Positions are not layout IDs.

RoboDojo · leaderboard comparisonRoboDojo · 排行榜方法对比

Selected leading entries, agent methods, and the GPT-6-Astra baseline. Auriga uses the current three-seed evaluation with all 6,300 episodes in the denominator.

选取领先条目、agent 方法以及 GPT-6-Astra 基线。Auriga 使用当前三 seed 评测，分母包含全部 6300 局。

Dimension-macro success rate维度等权成功率

RoboDojo selected methods: dimension-macro success rate / RoboDojo 所选方法维度等权成功率0%20%40%60%80%100%AurigaAuriga: 57.68%57.68%GO-3 Nano PreviewGO-3 Nano Preview: 56.42%56.42%PyRUA-LeanPyRUA-Lean: 50.35%50.35%HexaShell-PreviewHexaShell-Preview: 46.85%46.85%WorldScape-R²SIWorldScape-R²SI: 39.35%39.35%EmbodiedEvoEmbodiedEvo: 38.95%38.95%WALL-OSS-1.0-liteWALL-OSS-1.0-lite: 34.97%34.97%PhysicalRSIPhysicalRSI: 31.38%31.38%GPT-6-AstraGPT-6-Astra: 22.48%22.48%
Five dimensions weighted equally · success rate (%)五个能力维度等权 · 成功率（%）
All selected methods · dimension breakdown所选方法 · 各维度结果
Method方法Macro SR等权成功率Score / 1均分 / 1Generalization泛化Precision精细操作Memory记忆Long-Horizon长程Open开放指令
Auriga57.68%0.64048.89%38.50%92.78%45.17%63.08%
GO-3 Nano Preview56.42%0.61933.66%76.00%63.67%53.50%55.25%
PyRUA-Lean50.35%0.55343.17%38.25%68.33%43.75%58.25%
HexaShell-Preview46.85%0.53234.50%32.00%81.00%31.50%55.25%
WorldScape-R²SI39.35%0.46635.83%15.50%63.67%34.25%47.50%
EmbodiedEvo38.95%0.45130.83%16.25%76.67%21.75%49.25%
WALL-OSS-1.0-lite34.97%0.42637.94%29.08%56.00%32.25%19.58%
PhysicalRSI31.38%0.36315.62%32.50%46.56%37.33%24.92%
GPT-6-Astra22.48%0.29030.50%4.00%38.67%8.25%31.00%

Source: official RoboDojo leaderboard, 2026-10-30 snapshot, retrieved 2026-10-31. Baseline values are published leaderboard results; Auriga is our local evaluation. Generalization averages the standard and random splits; scores are rescaled to 0–1.

来源：RoboDojo 官方排行榜，2026-10-30 快照，2026-10-31 获取。基线取官网公布结果，Auriga 为本地评测。泛化维度对 standard 与 random 等权平均，分数统一换算为 0–1。

Cross-benchmark evaluation跨基准评测

RoboCasa365

Mobile manipulation across unseen kitchens. The same harness coordinates the arm, gripper, torso, and mobile base through action programs, without a learned VLA policy or cross-episode memory.

面向未见厨房环境的移动操作。同一套 harness 通过动作程序协调机械臂、夹爪、躯干与移动底盘，不使用学习式 VLA 策略或跨回合记忆。

73.2%overall success · 183 / 250 episodes总体成功率 · 183 / 250 局
62.5%composite tasks · 100 / 160 episodes组合任务成功率 · 100 / 160 局
Evaluation评测集
Target50 · 50 tasks · seeds 1–5 · 250 episodesTarget50 · 50 个任务 · seed 1–5 · 250 局
Environment环境
10 unseen target kitchens · Franka Panda arm on an Omron mobile base10 个未见目标厨房 · Franka Panda 机械臂与 Omron 移动底盘
Task groups任务分组
18 Atomic-Seen, 16 Composite-Seen, 16 Composite-Unseen tasks; chains of up to 15 subtasks18 个 Atomic-Seen、16 个 Composite-Seen、16 个 Composite-Unseen 任务；最长包含 15 个子任务
Model · budget模型 · 预算
GPT-6 astra · high · up to 100 model requests per episodeGPT-6 astra · high · 每局最多 100 次模型请求

Against RPent with the same model and reasoning effort: +14.0 percentage points overall and +19.4 points on composite tasks.

相较相同模型与推理档位的 RPent，总体成功率高 14.0 个百分点，组合任务高 19.4 个百分点。

Success across task groups各任务组成功率

AurigaRPentPyRUA-LeanRLDX-1
RoboCasa365 Target50 success rates in target kitchens / RoboCasa365 目标厨房各组成功率0%20%40%60%80%100%Overall总体Auriga · Overall: 73.2%73.2RPent · Overall: 59.2%PyRUA-Lean · Overall: 56.4%RLDX-1 · Overall: 30.0%Atomic-Seen原子 · SeenAuriga · Atomic-Seen: 92.2%92.2RPent · Atomic-Seen: 87.8%PyRUA-Lean · Atomic-Seen: 86.7%RLDX-1 · Atomic-Seen: 60.0%Composite All组合 · 合计Auriga · Composite All: 62.5%62.5RPent · Composite All: 43.1%PyRUA-Lean · Composite All: 39.4%RLDX-1 · Composite All: 13.1%Composite-Seen组合 · SeenAuriga · Composite-Seen: 67.5%67.5RPent · Composite-Seen: 43.8%PyRUA-Lean: not reported—RLDX-1 · Composite-Seen: 21.3%Composite-Unseen组合 · UnseenAuriga · Composite-Unseen: 57.5%57.5RPent · Composite-Unseen: 42.5%PyRUA-Lean: not reported—RLDX-1 · Composite-Unseen: 5.0%
Target kitchens · success rate (%) · — not reported目标厨房 · 成功率（%）· — 未报告
Method方法Execution执行方式Overall总体Atomic-Seen原子 · SeenComposite All组合 · 合计Composite-Seen组合 · SeenComposite-Unseen组合 · Unseen
AurigaAgent73.2%92.2%62.5%67.5%57.5%
RPentAgent + VLA (RLDX-1)59.2%87.8%43.1%43.8%42.5%
PyRUA-LeanAgent + VLA (RLDX-1)56.4%86.7%39.4%——
RLDX-1VLA30.0%60.0%13.1%21.3%5.0%

All three agent methods use GPT-6 astra at high reasoning effort. RPent and PyRUA-Lean delegate contact-rich actions to RLDX-1. RPent also uses task memory collected before evaluation. PyRUA-Lean reports composite performance only in aggregate; separate seen/unseen rates are not reported.

三个 agent 方法均使用 GPT-6 astra、high 推理。RPent 与 PyRUA-Lean 将接触密集动作交给 RLDX-1；RPent 还使用评测前收集的任务记忆。PyRUA-Lean 仅报告组合任务总体结果，未单独报告 seen/unseen 分组成功率。

Auriga · all 50 tasks and 250 episodesAuriga · 全部 50 个任务与 250 局结果
Task任务Group分组Seed 1Seed 2Seed 3Seed 4Seed 5Successes成功局数
CloseFridgeAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
CloseToasterOvenDoorAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
NavigateKitchenAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
OpenCabinetAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
OpenDrawerAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
PickPlaceCounterToStoveAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
PickPlaceDrawerToCounterAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
PickPlaceSinkToCounterAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
PickPlaceToasterToCounterAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
SlideDishwasherRackAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
TurnOffStoveAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
TurnOnElectricKettleAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
TurnOnMicrowaveAtomic-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
CloseBlenderLidAtomic-SeenSuccess成功Failure失败Success成功Success成功Success成功4/5
CoffeeSetupMugAtomic-SeenSuccess成功Failure失败Success成功Success成功Success成功4/5
OpenStandMixerHeadAtomic-SeenSuccess成功Success成功Failure失败Success成功Success成功4/5
PickPlaceCounterToCabinetAtomic-SeenSuccess成功Failure失败Success成功Failure失败Success成功3/5
TurnOnSinkFaucetAtomic-SeenSuccess成功Failure失败Failure失败Success成功Success成功3/5
DeliverStrawComposite-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
GetToastedBreadComposite-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
PreSoakPanComposite-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
RinseSinkBasinComposite-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
ScrubCuttingBoardComposite-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
WashLettuceComposite-SeenSuccess成功Success成功Success成功Success成功Success成功5/5
KettleBoilingComposite-SeenSuccess成功Success成功Success成功Failure失败Success成功4/5
PrepareCoffeeComposite-SeenFailure失败Success成功Success成功Success成功Success成功4/5
SetUpCuttingStationComposite-SeenFailure失败Success成功Success成功Success成功Success成功4/5
LoadDishwasherComposite-SeenSuccess成功Success成功Success成功Failure失败Failure失败3/5
SearingMeatComposite-SeenSuccess成功Success成功Success成功Failure失败Failure失败3/5
StackBowlsCabinetComposite-SeenSuccess成功Failure失败Success成功Success成功Failure失败3/5
StirVegetablesComposite-SeenFailure失败Failure失败Success成功Failure失败Success成功2/5
SteamInMicrowaveComposite-SeenFailure失败Failure失败Success成功Failure失败Failure失败1/5
PackIdenticalLunchesComposite-SeenFailure失败Failure失败Failure失败Failure失败Failure失败0/5
StoreLeftoversInBowlComposite-SeenFailure失败Failure失败Failure失败Failure失败Failure失败0/5
ArrangeTeaComposite-UnseenSuccess成功Success成功Success成功Success成功Success成功5/5
CuttingToolSelectionComposite-UnseenSuccess成功Success成功Success成功Success成功Success成功5/5
PortionHotDogsComposite-UnseenSuccess成功Success成功Success成功Success成功Success成功5/5
ArrangeBreadBasketComposite-UnseenSuccess成功Success成功Success成功Success成功Failure失败4/5
BreadSelectionComposite-UnseenSuccess成功Success成功Success成功Success成功Failure失败4/5
PanTransferComposite-UnseenSuccess成功Success成功Success成功Success成功Failure失败4/5
WeighIngredientsComposite-UnseenFailure失败Success成功Success成功Success成功Success成功4/5
GatherTablewareComposite-UnseenSuccess成功Success成功Success成功Failure失败Failure失败3/5
RecycleBottlesByTypeComposite-UnseenSuccess成功Failure失败Success成功Failure失败Success成功3/5
CategorizeCondimentsComposite-UnseenFailure失败Failure失败Success成功Failure失败Success成功2/5
HeatKebabSandwichComposite-UnseenFailure失败Success成功Success成功Failure失败Failure失败2/5
WaffleReheatComposite-UnseenFailure失败Success成功Failure失败Success成功Failure失败2/5
WashFruitColanderComposite-UnseenFailure失败Failure失败Success成功Success成功Failure失败2/5
GarnishPancakeComposite-UnseenSuccess成功Failure失败Failure失败Failure失败Failure失败1/5
MakeIceLemonadeComposite-UnseenFailure失败Failure失败Failure失败Failure失败Failure失败0/5
SeparateFreezerRackComposite-UnseenFailure失败Failure失败Failure失败Failure失败Failure失败0/5
Reference results · pretraining kitchens参考结果 · 预训练厨房

These reported policy results use the pretraining kitchens, not the unseen target kitchens above. They are listed separately because the evaluation environments differ.

这些策略结果在预训练厨房中评测，与上方未见目标厨房的环境不同，因此单独列出。

Method方法Overall总体Atomic-Seen原子 · SeenComposite All组合 · 合计Composite-Seen组合 · SeenComposite-Unseen组合 · Unseen
Xiaomi-Robotics-157.4%80.2%44.6%57.1%32.1%
WorldDreamer35.3%66.3%17.9%26.7%9.0%
π₀.₅16.9%39.6%4.2%7.1%1.2%
π₀14.8%34.6%3.6%6.1%1.1%

Source: Auriga technical report, Experimental Setup and Main Results, Table “RoboCasa365 Target50” (2026-10-31 source snapshot). Auriga counts are checked against all 250 per-episode records; baseline values follow the report.

来源：Auriga 技术报告的实验设置、主要结果及 RoboCasa365 Target50 表（2026-10-31 源码快照）。Auriga 统计已核对全部 250 条逐局记录，基线数值沿用报告。

Architecture架构

Decision, capability, admission决策、能力与执行准入

Cortex maintains the task belief and chooses each action. General tools provide measurements and motion previews; one admission gateway executes the previewed trajectory, checks each movement stage, and returns evidence for the next decision.

Cortex 维护任务信念并选择动作。通用工具提供测量与运动预览；所有物理动作经过同一准入网关，执行预览轨迹、逐阶段验证，并返回证据供模型继续决策。

Auriga: Cortex, capability layer, admission gateway, task belief and evidence record HARNESS / RUN CONTROL 框架 / 运行控制 Task instruction + initial observation 任务指令 + 初始观察 Cortex Cortex Persistent model session · decides and recovers 持续模型会话 · 决策与恢复 Task belief B: Plan / Attention / Ledger 任务信念 B：计划 / 关注项 / 前提账本 Goals, uncertainties, and evidence-backed premises 目标、不确定性，以及有证据支持的前提 Capability layer 能力层 Measure · Act · Compute · Recall 测量 · 动作 · 计算 · 检索 Admission gateway 执行准入网关 Preview → cached execution → stage verification 预览 → 执行缓存轨迹 → 逐阶段验证 Tolerance stops · model-placed checkpoints 超出容差即停止 · 模型设置观察检查点 Tool calls / results 工具调用 / 结果 Physical requests / receipts 物理动作请求 / 执行回执 Evidence record E 证据记录 E Requests · receipts · views · measured trajectories 动作请求 · 回执 · 图像 · 实测轨迹 Recall 检索 Robot environment 机器人环境 RoboDojo / RoboCasa Actions / observations 动作 / 观察 Web workbench Web 工作台 Inspect · archive · replay 查看 · 归档 · 回放
Direct and program-generated motions share the same gateway. Dashed paths store and retrieve evidence; task belief remains the model’s revisable interpretation. The workbench supports inspection and replay.直接动作与程序生成的动作共用准入网关。虚线路径用于记录和检索证据；任务信念是模型可修订的判断。工作台提供查看与回放。
Run loop: observe, plan, preview, execute, feedback, then continue or stop ObserveBeliefPreview ExecuteFeedback Done? continue / handle failure adjust stop → scoring, archive, replay
STEP 01

Observe & measure

Calibrated RGB views and measured robot state ground the next decision. Tools add 3D object estimates and geometric measurements with explicit reliability indicators.

Observation, measurement, retrieval, and preview leave the simulated world paused.

步骤 01

观察与测量

标定的 RGB 视图与实测机器人状态为决策提供依据；工具补充带有可靠性说明的物体三维估计与几何测量。

观察、测量、检索与预览均不推进仿真时间。

STEP 02

Maintain task belief

Before the first physical action, Cortex submits a task belief: a plan, an attention list of uncertainties, and a ledger of task-critical premises.

Premises remain pending, supported, contradicted, or unknown according to evidence. The latest belief survives context compaction; detailed evidence stays retrievable.

步骤 02

维护任务信念

首次物理动作前，Cortex 提交任务信念：计划、记录不确定性的关注列表，以及记录关键前提的账本。

前提根据证据标记为待验证、有支持、被反驳或未知。上下文压缩后保留最新信念，详细证据仍可检索。

STEP 03

Preview before acting

Every requested motion, including one generated by a program, is compiled into a trajectory preview with paths and stage timing before it can execute.

The model reviews the preview by default. It may skip a separate review turn, but execution still uses the cached previewed trajectory without replanning.

步骤 03

执行前预览

每个运动请求，包括程序生成的动作，都先编译为包含路径与阶段时长的轨迹预览。

模型默认检查预览，也可跳过单独的检查轮次；执行仍使用已预览的缓存轨迹，不重新规划。

STEP 04

Execute & verify

The gateway executes the cached trajectory and checks the reached pose at the end of each movement stage. A deviation beyond the model’s declared tolerance stops the remaining stages.

Model-placed checkpoints pause for fresh observations, optionally triggered by gripper aperture. Completed motion is not undone.

步骤 04

执行与验证

网关执行缓存轨迹，在每个运动阶段结束时检查到达位姿。偏差超出模型声明的容差时，停止后续阶段。

模型设置的检查点会暂停执行并返回新观察，也可由夹爪开度条件触发；已完成的运动不会撤销。

STEP 05

Assess & revise

The receipt reports completed and unexecuted stages, measured deviations, and fresh views. Cortex diagnoses the outcome and chooses whether to continue, gather evidence, or preview a revised motion.

Outcomes update the task belief and evidence record. The benchmark evaluator determines task success independently of motion completion or normal program termination.

步骤 05

评估与修正

回执报告已完成与未执行阶段、实测偏差和新视图。Cortex 诊断结果，决定继续、收集证据，或预览修订后的动作。

动作结果更新任务信念与证据记录。任务是否成功由基准评估器独立判断，区别于运动完成或程序正常退出。

Design principles设计原则

Four commitments四条原则

The model decides what to do, tools provide the capabilities, and the harness owns the boundary between decision and physical execution.

由模型决定做什么，工具提供能力，框架负责决策与物理执行之间的边界。

  1. 01

    Explicit task belief显式任务信念

    Keep goals, unresolved questions, and supported premises in a compact working state; check them against retrievable evidence.

    将目标、待解决问题和有支持的前提保存在精简的工作状态中，并通过可检索证据核对。

  2. 02

    Preview is the execution contract预览约束执行

    Direct and program-generated motions execute the cached previewed trajectory. Inspection changes no simulated state.

    直接动作与程序生成的动作执行已预览的缓存轨迹，预览检查不改变仿真状态。

  3. 03

    Verify each movement stage逐阶段验证

    Declared tolerances and selective observation checkpoints expose errors before later stages compound them.

    通过声明的容差与按需设置的观察检查点，及时暴露误差，避免后续阶段放大问题。

  4. 04

    The model chooses recovery模型决定恢复方式

    The gateway stops and reports; Cortex interprets the evidence and changes the action or task belief.

    网关负责停止和报告，Cortex 解释证据，调整动作或任务信念。

Control paradigms控制范式

Three ways to put a model on a robot让模型控制机器人的三种方式

The paradigms differ in who decides, what reaches the actuators, and how failures are noticed and repaired.

三种范式的区别在于：由谁决策、什么指令到达执行器，以及如何发现和修复失败。

  (a)Action model (VLA)动作模型（VLA） (b)General model, direct control通用模型直接控制 (c) · ours本方法Auriga
Decides决策 A learned policy maps images and the instruction to actions.由学习得到的策略将图像和指令映射为动作。 A general model chooses the next action at each step (e.g. GPT Direct).通用模型在每一步选择下一个动作（如 GPT Direct）。 A persistent Cortex session maintains task belief and retrieves episode evidence.持续的 Cortex 会话维护任务信念，并检索本局证据。
Executes执行 Predicted action chunks go straight to the robot.预测的动作块直接下发给机器人。 One tool call, one robot step.一次工具调用对应机器人的一步动作。 Direct or program-generated motions use one gateway and the cached previewed trajectory.直接动作或程序生成的动作通过同一网关，执行已预览的缓存轨迹。
Feedback反馈 Implicit, through the next camera frame.隐式反馈，通过下一帧相机图像获得。 A new observation after each call.每次调用后获得一次新的观察。 Fresh views, pose deviations, completed stages, and remaining stages in each execution receipt.执行回执包含新视图、位姿偏差、已完成阶段与剩余阶段。
Recovery恢复 Bounded by what the training data covers.受限于训练数据的覆盖范围。 Re-decided at each call from the new frame; no plan preview or execution receipt.每次调用根据新图像重新决策；没有计划预览和执行回执。 The gateway stops on failed checks; Cortex diagnoses, resumes, or revises its action and belief.检查失败时网关停止，Cortex 负责诊断、继续，或修订动作与信念。
Tool layer工具层

Measure, act, compute, recall测量、动作、计算与检索

General capabilities turn observations into measurements with explicit reliability and model decisions into reviewable actions. Deliberation leaves the simulated world paused; only physical requests advance it through the admission gateway.

通用能力将观察转化为带有可靠性说明的测量，将模型决策转化为可检查的动作。思考与查询期间仿真保持暂停，只有经过准入网关的物理请求才推进世界。

Click a tool to see what it does, what goes in and out, and an illustrative workflow.点击工具查看用途、输入输出与使用流程。

Citation文献引用

Cite this work引用

StellarEdge Team. Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation. Project website, 2026. https://auriga.stelledge.com/.

StellarEdge Team. Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation. 项目网页，2026。https://auriga.stelledge.com/。

@misc{auriga2026,
  author       = {{StellarEdge Team}},
  title        = {{Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation}},
  year         = {2026},
  howpublished = {Project website},
  url          = {https://auriga.stelledge.com/}
}