Company in stealth · public brief only公司處於 stealth 階段 · 僅公開概要ed.wei@robotsys.ai
RobotSys.ai
EMPOWERING A NEW CIVILIZATION WITH SILICON CARRIERS
Robots can already move. Nobody can yet say whether they can be trusted.機器人已經會動了。還沒有人能說清楚，它能不能被信任。
Every technical revolution needed a unit of measure before it scaled. Steam had horsepower. Electricity had volts and amperes. Semiconductors had yield and PPA. Embodied AI still has none.每一次技術革命，都要先有一套量測標準才真正起飛。蒸汽機有馬力，電力有伏特與安培，半導體有良率與 PPA。具身 AI 現在還沒有。
RobotSys.ai is operating in stealth. This site carries the industry-level material only: the bottlenecks that hold autonomous robots back, and the shape of the method we use to measure them. Everything specific to the company is released in a briefing.RobotSys.ai 目前處於 stealth 階段。本站只放產業層級的內容：自主機器人的核心瓶頸，以及我們用來量測它們的方法輪廓。公司自身的資料在說明會中提供。
Public here此處公開
The sixteen core bottlenecks, in full, with figures十六項核心瓶頸的完整說明與圖示
The four conditions safety requires at once安全必須同時成立的四個條件
What differentiation testing is, and why a pass/fail result is not enough鑑別測試是什麼，以及為何通過與否並不足夠
The three levels we measure at, in outline三個量測層級的輪廓
Released in a briefing說明會中提供
Platform architecture, operating loop and unit economics平台架構、營運循環與單位經濟
Test contracts, metric definitions and reference sets測試合約、指標定義與參考集
Partners, facilities and the first commercial platform合作夥伴、場域與第一個商業平台
Roadmap, team and funding plan時程、團隊與募資計畫
THE FUNDAMENTAL QUESTION
Safety requires four conditions at once安全需要四個條件同時成立
The state has to be known, the action has to be acceptable, the physical effect has to arrive in time, and the whole claim has to keep holding. No existing method checks all four together.狀態要知道、動作要能被接受、物理效果要來得及，而且這個結論要能一直成立。今天沒有任何一套現成方法能同時檢驗這四件事。
2.01 — 2.16
Sixteen bottlenecks, all of them public十六項瓶頸，全部公開
These problems share one property: none of them can be answered with a pass or a fail. They only become visible when several robots are placed under the same set of conditions and those conditions are allowed to move.這些問題有一個共同點：都無法用「有沒有通過」來回答。它們只有在多台機器人被放進同一組會變動的條件時，才會顯形。
Select any tile for the full entry, the test that exposes it, and the technical view.點任一格進入該項的完整說明、鑑別方式與技術視角。
WHAT WE DO
Two things, and they reinforce each other我們做兩件事，而且互相支撐
01
Differentiation testing鑑別測試
A standard evaluation confirms whether a robot met the bar. A differentiation test moves the conditions inside the application's own envelope until the machines separate, then reports who stays stable, who degrades first, and where the gap opens.一般評測確認有沒有達標。鑑別測試在應用合理範圍內調整條件，直到機器彼此分開，然後回報誰還穩定、誰先退化、差距從哪裡開始出現。
02
Bottleneck localisation應用瓶頸定位
Along the closed loop from observation to the next observation, we locate which segment is actually limiting performance, using external behaviour and test conditions rather than the vendor's internal implementation.沿著從觀測回到下一次觀測的閉迴路，定位真正限制效能的是哪一段。依據外部行為與測試條件，而非廠商的內部實作。
Turning the differences between several robots into engineering knowledge someone can act on.把多台機器人的差異，轉成可行動的工程知識。
STAY IN THE LOOP
Everyone sees the demo. Almost nobody sees the retest.所有人都看得到那支示範影片。幾乎沒有人看得到複測。
We spend our weeks putting robots into conditions their launch videos avoid. What we learn goes out to this list first — in plain numbers, with the conditions attached.我們每一週都在把機器人放進發表影片會避開的條件裡。學到的東西會先寄給這份名單——用實際數字，並且附上測試條件。
Where autonomous-robot capability actually stands, and what genuinely moved this month自主機器人的能力現況，以及這個月真正有進展的部分
Market and supply-chain developments worth an engineer's attention值得工程端注意的市場與供應鏈動態
New releases from us, teaching videos, and how to actually use them我們的新產品發布、教學影片，以及實際怎麼使用
Events we are speaking at or hosting, with enough notice to book a seat我們參與或舉辦的活動，提前到還來得及報名
NO FIXED SCHEDULE · ONLY WHEN THERE IS SOMETHING WORTH SENDING沒有固定週期 · 有值得寄的東西才寄
Your address is stored only to send these updates. One click unsubscribes, and we never pass it on.您的信箱只用於寄送這些更新。隨時可退訂，我們不會提供給第三方。
HIRING
We are still building the team團隊還在擴編
Physical AI, world models, autonomous systems, robotic systems and electromechanical design — five areas, open now.Physical AI、World Model、自主系統、機器人系統、機電系統設計——五個方向，現正招募。
The sixteen core bottlenecks of autonomous robots自主機器人的十六項核心瓶頸
Each entry carries a scenario figure and a technical view, is tagged with the safety condition it threatens, and names the level at which we measure it.每一項都附上情境圖與技術視角圖，標註它威脅的是哪一個安全條件，以及由哪一層負責量測。
Select a tag to filter. Select any figure to enlarge.點按標籤可篩選。點任一張圖可放大檢視。
2.01KNOWLevel 3
The current state may be unknown現在的狀態可能是未知的
現在的狀態可能是未知的The current state may be unknown
Unseen space can turn a valid plan into unsafe motion沒被看到的空間，會讓一條合法的路徑變成危險的動作沒被看到的空間，會讓一條合法的路徑變成危險的動作Unseen space can turn a valid plan into unsafe motion
Why it is hard難在哪裡
Sensing is occluded, data goes stale, estimates drift. Treating “not observed” as “not present” produces a path that is perfectly valid on the map and collides with a toy on the floor.感測有遮蔽、資料有延遲、估測會漂移。機器人把「沒觀測到」當成「沒有東西」，於是規劃出一條在地圖上完全合法、在現實中會撞上地板玩具的路徑。
How a differentiation test exposes it鑑別測試怎麼看這一條
Place occlusions and low-profile obstacles in the field, then observe how each robot treats unobserved space — slow down and probe, or assume clear. Across several machines this separates immediately.在測試場刻意佈置遮蔽區與低矮障礙，觀察機器人對未觀測區的處理方式——減速探查，或直接視為淨空。多台同場測試時，差異在這裡立刻分開。
Technical view技術視角2.01
The state estimate comes from observation, and observation is degraded by occlusion, staleness and drift.估測狀態來自觀測，觀測本身受遮蔽、資料過期與雜訊漂移三重影響。
2.02ACTLevel 1
Reaction time is the time to useful physical effect反應時間算到物理效果發生的那一刻
反應時間算到物理效果發生的那一刻Reaction time is the time to useful physical effect
Reaction time ends when the world is physically changed反應時間結束在世界被物理改變的那一刻反應時間結束在世界被物理改變的那一刻Reaction time ends when the world is physically changed
Why it is hard難在哪裡
From detecting the slip, to updating the command, to torque building, to the arm moving, to the glass actually being caught — every segment costs time. Command latency can look excellent while the interception window has already closed.從偵測到滑落、更新指令、馬達建立扭力、手臂移動，到真的接住杯子，每一段都要花時間。指令延遲的數字很漂亮，攔截窗口還是可能已經過去。
How a differentiation test exposes it鑑別測試怎麼看這一條
Measure end-to-end latency with a repeatable stimulus–response fixture that terminates at a physical quantity — displacement or force — rather than at command issue.用可重複的刺激—反應治具量測端到端延遲，量到物理量（位移、力），而非量到指令送出。
Technical view技術視角2.02
Sensor delay, actuator dynamics and object momentum all sit inside the loop. A fast code update is not a fast physical response.感測延遲、致動器動態與物件動量都落在迴路裡。程式更新得快，物理回應不會跟著快。
2.03ACTLevel 1 / 2
Physical authority can arrive too late物理權限可能來不及
物理權限可能來不及Physical authority can arrive too late
Control authority must arrive before the remaining safety margin is gone控制權限必須在剩餘安全餘裕消失之前抵達控制權限必須在剩餘安全餘裕消失之前抵達Control authority must arrive before the remaining safety margin is gone
Why it is hard難在哪裡
The brake command is correct. Payload momentum, floor traction and the maximum braking force available at the current battery and thermal state together set the real stopping distance. The safety line is still crossed.煞車指令完全正確，但載重動量、地面摩擦，以及電量與溫度限制下能給出的最大制動力，共同決定實際停止距離。安全線還是會被跨過。
How a differentiation test exposes it鑑別測試怎麼看這一條
Measure the gap between required and available stopping distance under low traction and eccentric load, then sweep charge and temperature to plot how braking authority degrades with condition.在低摩擦地面與偏心負載條件下，量測實際停止距離與所需距離的差，並掃描電量與溫度區間，畫出制動權限隨條件退化的曲線。
Technical view技術視角2.03
Detection is timely, the decision is right, the command is sent — and the physical effect still lands outside the safety window.偵測及時、決策正確、指令送出，物理效果仍可能落在安全窗口之外。
2.04ACTLevel 2
Precision, compliance and impact tolerance conflict精度、柔順與耐衝擊互相衝突
精度、柔順與耐衝擊互相衝突Precision, compliance and impact tolerance conflict
Better impact tolerance can reduce immediate positioning certainty更好的耐衝擊會降低當下的定位確定性更好的耐衝擊會降低當下的定位確定性Better impact tolerance can reduce immediate positioning certainty
Why it is hard難在哪裡
A rigid wrist aligns precisely and transmits a harsh contact force. A compliant wrist absorbs the impact and shifts position. The task needs both, and physics offers no free maximum.剛性手腕對得準，接觸瞬間的力也大。柔順手腕吸收得掉衝擊，位置就會偏。任務同時需要這兩件事，物理上沒有免費的最大值。
How a differentiation test exposes it鑑別測試怎麼看這一條
On a single insertion task, measure the trade-off curve between alignment error and peak contact force. That curve is the action capability fingerprint of the mechanism.在同一個插接任務上，量測對位誤差與接觸峰值力的取捨曲線。這條曲線本身就是該機構的 action capability fingerprint。
Technical view技術視角2.04
Rigidity buys repeatability at the cost of harsher contact; compliance absorbs energy but needs travel and control margin.剛性帶來可重複性與較劇烈的接觸；柔順吸收能量，需要行程與控制餘裕。
2.05SUSTAINLevel 1
Average timing does not establish safety平均反應時間不能證明安全
平均反應時間不能證明安全Average timing does not establish safety
Safety depends on bounded worst-case timing, not the average安全取決於有界的最壞情況時序安全取決於有界的最壞情況時序Safety depends on bounded worst-case timing, not the average
Why it is hard難在哪裡
A system averaging 100 ms looks fast. One 750 ms tail event that coincides with a hand entering the work zone is enough to invalidate the safety argument. Once is enough.平均 100 毫秒的系統看起來很快，尾端那次 750 毫秒的延遲只要落在手伸進作業區的時刻，安全論證就不成立。一次就夠。
How a differentiation test exposes it鑑別測試怎麼看這一條
Trigger repeatedly over long runs, collect the full latency distribution and read P99.9 and the maximum. Inject compute, network and thermal load to see how the tail opens up.長時間重複觸發，收集完整延遲分布，看 P99.9 與最大值。同時注入運算、網路與熱負載，觀察尾端如何被拉開。
Technical view技術視角2.05
The average sits in the head of the distribution; the safety deadline sits in the tail.平均值落在分布的頭部，安全期限落在尾部。要證明的是尾端有界。
2.06DECIDELevel 2 / 3
Local correctness can create whole-body failure局部正確會造成整機失效
局部正確會造成整機失效Local correctness can create whole-body failure
Local task success does not prove whole-body safety局部任務成功，不能證明整機安全局部任務成功，不能證明整機安全Local task success does not prove whole-body safety
Why it is hard難在哪裡
The base reaches its path target and the arm reaches its reach target. Both controllers report success. Together, the centre of mass has left the support polygon.底盤走到指定路徑、手臂到達指定位置，兩個控制器各自回報成功。合起來，質心已經跑出支撐多邊形。
How a differentiation test exposes it鑑別測試怎麼看這一條
Command coordinated base and reach motion, then measure stability margin, support-polygon reserve and the tip-over boundary to find where each part is correct and the whole is not.同時下達移動與伸展的組合指令，量測穩定裕度、支撐多邊形餘量與傾覆邊界，找出兩個子系統各自正確、合起來越界的條件。
Technical view技術視角2.06
Vision, legs and arm share one body. Balance, clearance, force and energy are shared limits that combined motion consumes at once.視覺、腿、手臂共用同一個身體。平衡、間隙、力量與能量是共用限制，組合動作會同時消耗它們。
2.07DECIDELevel 3
A safe action may not exist, and a recovery action may also be unsafe安全動作可能不存在，回復動作也可能不安全
安全動作可能不存在，回復動作也可能不安全A safe action may not exist, and a recovery action may also be unsafe
A corrective action is not automatically a safe action修正動作不會自動成為安全動作修正動作不會自動成為安全動作A corrective action is not automatically a safe action
Why it is hard難在哪裡
Front clearance is shrinking, rear space is limited, and stopping leaves the payload momentum in place. Every option carries risk, and the safe set may already be empty.前方淨空在縮小、後方空間有限、停下來載重的動量還在。三個選項都帶著風險，安全集合可能已經是空的。
How a differentiation test exposes it鑑別測試怎麼看這一條
Narrow the scenario to tight corridors under time pressure and observe what the robot chooses as the safe set approaches empty — and whether it knows it has run out of safe options.把場景收斂到窄通道與時間壓力，觀察機器人在安全集合接近空集時的選擇，以及它是否知道自己已經沒有安全選項。
Technical view技術視角2.07
The safe set in state space shrinks over time and can become empty. Any recovery then has to be assessed again.狀態空間中的安全集合會隨時間縮小，太晚就會成為空集。此時任何回復動作都必須重新評估。
2.08KNOWLevel 2
The robot may not know its own current capability機器人可能不知道自己現在的能力
機器人可能不知道自己現在的能力The robot may not know its own current capability
Safety requires current capability, not yesterday’s calibration安全取決於當下的能力，而非昨天的校正安全取決於當下的能力，而非昨天的校正Safety requires current capability, not yesterday’s calibration
Why it is hard難在哪裡
Battery voltage sags, motors run hot, the payload changed. The holding force actually available is now below what the plan assumed, and the plan is still running on expected capability.電池電壓下降、馬達過熱、負載改變，實際可用的握持力已經低於計畫假設的值。規劃仍然照著預期能力在走。
How a differentiation test exposes it鑑別測試怎麼看這一條
Re-run the same actions at low charge, high temperature and worn condition, compare declared against present capability, and check whether the system degrades itself.在低電量、高溫與磨耗條件下重測同一組動作，比較「宣稱能力」與「當下能力」的落差，並檢查系統是否會自己降級。
Technical view技術視角2.08
The force and braking envelope contracts with charge, temperature and wear, while the plan still quotes the outer curve.力量／制動能力對速度的包絡線會隨電量、溫度與磨耗整體內縮，計畫卻仍然引用外圈的那條線。
2.09KNOWLevel 2 / 3
Uncertainty cannot be completely learned away不確定性無法完全被學掉
不確定性無法完全被學掉Uncertainty cannot be completely learned away
Learning reduces uncertainty; it does not reveal the unseen學習會縮小不確定性，不會揭露看不見的東西學習會縮小不確定性，不會揭露看不見的東西Learning reduces uncertainty; it does not reveal the unseen
Why it is hard難在哪裡
What is inside the bag, how the weight is distributed, where the fragile item sits, how friction varies — these hidden states do not appear because there is more training data. High confidence and certainty are different things.袋子裡的內容物、重量分布、易碎品位置、變動的摩擦，這些隱藏狀態不會因為訓練資料變多就自己出現。高信心與確定性是兩件事。
How a differentiation test exposes it鑑別測試怎麼看這一條
Use test articles with controlled hidden state — identical appearance, different contents — and measure whether the policy probes conservatively or commits to its training distribution.用隱藏狀態受控的測試件（同外觀、不同內容），量測策略在未知條件下的行為——保守探查，或直接依訓練分布下手。
Technical view技術視角2.09
More data narrows the distribution. Residual uncertainty remains; confidence rises and the unknown does not disappear.更多資料會把分布收窄，殘餘不確定性仍然存在。信心分數升高，未知並沒有消失。
2.10ACTLevel 2
Contact is discontinuous, and many household objects are not rigid接觸是不連續的，很多居家物件不是剛體
接觸是不連續的，很多居家物件不是剛體Contact is discontinuous, and many household objects are not rigid
Contact is an event, not a smooth continuation接觸是一個事件，而不是平滑的延續接觸是一個事件，而不是平滑的延續Contact is an event, not a smooth continuation
Why it is hard難在哪裡
Cloth snags, force builds, and at the moment of release force collapses and the arm overshoots. Smooth predictive models fail exactly at that instant.布料被勾住、力量持續上升、鬆脫的瞬間力量瞬降，手臂衝過頭。平滑的預測模型在這個瞬間直接失效。
How a differentiation test exposes it鑑別測試怎麼看這一條
Create repeatable snag-and-release events with a fixture, measure overshoot and recovery time after the force discontinuity, and compare how different robots handle event-type contact.用可重複的勾住—鬆脫治具製造接觸事件，量測力量突變後的過衝量與恢復時間，比較不同機器人對事件型接觸的處理。
Technical view技術視角2.10
Deformation, slip and mass redistribution change force within milliseconds. A small state error can produce a large force change.變形、滑移與質量重分布都會讓力量在極短時間內改變。很小的狀態誤差可以造成很大的力量變化。
2.11DECIDELevel 3
Physical safety does not imply human-acceptable safety物理安全不等於人可以接受的安全
物理安全不等於人可以接受的安全Physical safety does not imply human-acceptable safety
No contact does not mean consent沒有碰撞，不等於得到同意沒有碰撞，不等於得到同意No contact does not mean consent
Why it is hard難在哪裡
No collision, low speed, clear path — physically safe. It also captured what was on your screen, or moved inside a distance you did not want it to enter.沒有碰撞、速度很低、路徑淨空，物理上完全安全。它同時把你螢幕上的內容拍了進去，或站進了你不想被靠近的距離。
How a differentiation test exposes it鑑別測試怎麼看這一條
Write privacy boundaries, personal space and respect for intent into the acceptance conditions of the Application Profile, and score human intervention and acceptability rather than collisions alone.把隱私邊界、個人空間與意圖尊重寫進 Application Profile 的驗收條件，用人為介入次數與可接受度評分當輸出，而非只看碰撞次數。
Technical view技術視角2.11
Intent, privacy, hygiene and personal space all sit inside human-acceptable safety, and none of them follows from a force limit and a clear path.意圖、隱私、衛生與個人空間都在人可接受的安全範圍內，這四項無法由力量上限與淨空路徑推導出來。
Memory alone is not enough只有記憶並不夠只有記憶並不夠Memory alone is not enough
Why it is hard難在哪裡
The task was interrupted, the cup was moved, someone already watered the plant. The robot remembers the instruction and not what changed, so it repeats an action that is now wrong.任務被中斷、杯子被移動、有人已經澆過水。機器人記得原本的指令，卻不知道世界在這段時間改變了什麼，於是重複執行一個已經錯的動作。
How a differentiation test exposes it鑑別測試怎麼看這一條
Inject external change mid-task — move the target, let a person complete part of the subtask — and measure whether the robot re-verifies state, resumes safely, and what a repeated action costs.在任務中途注入外部改變（移動目標、由人完成部分子任務），量測機器人是否重新確認狀態、能否安全接續，以及重複執行造成的代價。
Technical view技術視角2.12
An old instruction becomes the wrong action once someone else has completed part of the task任務中途被人完成一部分，舊指令就成了錯的動作
Past actions must still explain the present. Once the causal chain breaks, every further decision rests on a stale premise.過去的動作必須仍然能解釋現在的狀態。因果鏈一旦斷開，續行的每一個決策都建立在過期的前提上。
2.13KNOWLevel 2 / 3
Physical-interaction data and generalization remain limited物理互動資料與泛化能力仍然有限
物理互動資料與泛化能力仍然有限Physical-interaction data and generalization remain limited
A learned skill may fail on a physically different object學過的技能可能在物理性質不同的物件上失敗學過的技能可能在物理性質不同的物件上失敗A learned skill may fail on a physically different object
Why it is hard難在哪裡
Two boxes look identical and differ threefold in weight. Visual similarity is not physical similarity, and training cannot cover every weight, friction, stiffness and load.外觀一樣的箱子，重量差三倍。視覺相似不等於物理相似。訓練資料涵蓋不到每一種重量、摩擦、剛性與負載組合。
How a differentiation test exposes it鑑別測試怎麼看這一條
Build a test set of objects with matched appearance and different physical properties, measure how transfer success decays, and observe whether the robot verifies physical response before committing force.用同外觀、不同物理性質的物件組成測試集，量測技能移轉的成功率衰減，並觀察機器人是否在出力前先驗證物理回應。
Technical view技術視角2.13
Contact data is expensive and sparsely covered. Success in training does not prove success in the field.接觸資料的取得成本高、覆蓋稀疏。訓練場的成功無法證明現場的成功。
2.14SUSTAINLevel 1–3
Learning and operational change can invalidate earlier safety evidence學習與運作條件的改變會讓既有安全證據失效
學習與運作條件的改變會讓既有安全證據失效Learning and operational change can invalidate earlier safety evidence
A passed test is not a permanent safety guarantee通過過的測試不是永久的安全保證通過過的測試不是永久的安全保證A passed test is not a permanent safety guarantee
Why it is hard難在哪裡
It passed at 0.5 m/s. After a software update the operating speed is 1.0 m/s. The same stop command now yields a different stopping distance, and the old evidence was bound to the old configuration.測試時 0.5 m/s 通過，軟體更新後運作速度變成 1.0 m/s。同一個停止指令，停止距離已經不同。舊證據綁定的是舊組態。
How a differentiation test exposes it鑑別測試怎麼看這一條
Bind test contracts to configuration versions, so a model update, hardware change or new operating condition triggers a retest quantified as a before/after result.把測試合約與組態版本綁在一起，模型更新、硬體更換或運作條件改變即觸發複測，並用 Before／After 量化差異。
Technical view技術視角2.14
Model updates, hardware wear and operating change all detach earlier evidence from the premises it was established under.模型更新、硬體磨耗與運作條件改變，都會讓早先的安全證據脫離它成立的前提。
2.15SUSTAINLevel 1 / 2
Product limits constrain every capability simultaneously產品限制同時壓縮所有能力
產品限制同時壓縮所有能力Product limits constrain every capability simultaneously
One product budget limits many capabilities at once一份產品預算同時限制多項能力一份產品預算同時限制多項能力One product budget limits many capabilities at once
Why it is hard難在哪裡
Size, weight, cost, power, cooling and noise share one budget. Sensing range, compute, actuation and runtime compete inside it. More of one is less of another.尺寸、重量、成本、功耗、散熱與噪音共用一份預算。感測範圍、算力、致動能力與續航互相爭奪。多給一項，就少給另一項。
How a differentiation test exposes it鑑別測試怎麼看這一條
Measure the trade-off surface across capabilities under a fixed SWaP-C budget so different architectures can be compared on equal terms, exposing architecture-level headroom.在同一個 SWaP-C 預算下量測各項能力的取捨面，讓不同架構在同一條件下比較，找出架構層級的改善空間。
Technical view技術視角2.15
SWaP-C is a shared budget, and it bounds every safety and performance claim made on top of it.SWaP-C 是共用預算。任何一項安全或效能宣稱，都受這份預算限制。
2.16SUSTAINLevel 1 / 2
Degradation is a safety problem, not only a maintenance problem退化是安全問題，不只是保養問題
退化是安全問題，不只是保養問題Degradation is a safety problem, not only a maintenance problem
A robot can still work and no longer be safe機器人可以還在運轉，同時已經不再安全機器人可以還在運轉，同時已經不再安全A robot can still work and no longer be safe
Why it is hard難在哪裡
Pads wear, friction drops, sensors drift, the battery ages. The robot still looks fully operational while most of the safety margin has already been consumed.夾墊磨損、摩擦下降、感測漂移、電池老化。機器人看起來一切正常，安全裕度已經被吃掉一大半。
How a differentiation test exposes it鑑別測試怎麼看這一條
Run accelerated wear and long-duration tests, track the safety margin over time, and separate “still working” from “still safe” as two independent verdicts.用加速磨耗與長時間運行測試，追蹤安全裕度隨時間的曲線，把「還能動」與「還安全」分成兩個獨立的判定。
Technical view技術視角2.16
Sensor drift, actuator wear, battery ageing and changing friction all narrow the margin. A maintenance signal becomes a safety signal.感測漂移、致動器磨耗、電池老化與摩擦改變會一路縮小安全裕度。保養訊號會變成安全訊號。
HOW THEY CHAIN
These sixteen are not independent這十六項不是各自獨立的
The world or the robot changes, the state becomes incomplete or stale, judgment and capability become uncertain, constraints turn wrong or conflicting, too little physical authority remains, and the outcome is unsafe or unacceptable. Safety needs continuity across the whole chain.世界或機器人改變、狀態變得不完整或過期、判斷與能力變得不確定、限制條件錯誤或互相衝突、剩下的物理權限太少，最後成為不安全或不可接受的結果。安全需要整條鏈上的連續性。
TECHNICAL OVERVIEW
One requirement, taken down three levels and brought back up一份需求，往下走三層，再走回來
A buyer states what has to be acceptable. That statement is decomposed until every piece of it is a loop small enough to measure repeatably, and the measurements are then carried back up into a claim about behaviour. Both directions matter: without the downward path a failure has no explanation, and without the upward path a measurement has no meaning.買方說出什麼樣的結果可以被接受。這句話被往下拆解，直到每一塊都成為小到可以重複量測的迴路；量測結果再往上帶回，成為關於行為的主張。兩個方向都要有：缺了向下這條路，失敗就沒有解釋；缺了向上這條路，量測就沒有意義。
Level層級
Subject under test待測物
What is varied變動的是什麼
Primary output主要輸出
Question it answers它回答的問題
Level 3 · behaviourLevel 3 · 行為
Complete robots, AMRs, mobile manipulators, humanoids, UAVs完整機器人、AMR、mobile manipulator、人形機器人、無人機
Scene, clutter, interruption, human presence, duration場景、雜亂度、中斷、有人在場、持續時間
Success rate, completion time, human intervention, cost of operation成功率、完成時間、人為介入、運作代價
Can this machine do the job in my building?這台機器能不能在我的場域完成工作
Level 2 · actionLevel 2 · 動作
Arms, grippers, mobile bases, joint-module assemblies機械臂、手爪、移動底盤、joint module 組合
Latency, jitter, settling time, overshoot, thermal drift延遲、抖動、穩定時間、過衝、熱漂移
Is the physical response bounded?物理回應是否有界
DIFFERENTIATION TESTING
The difference is not whether it passed差別不在有沒有過關
Under one fixed set of conditions, three capable machines all pass and the report says very little. Move the conditions inside what the application genuinely demands, and the same three separate.在同一組固定條件下，三台有能力的機器都會通過，報告也就說不出什麼。把條件調整到應用真正會出現的範圍，同樣這三台就分開了。
The standard test sits where every curve is still flat. The useful information sits to the right of it.一般評測落在所有曲線都還平坦的位置。有用的資訊落在它的右邊。
Once success rate sits above 99%, it stops being the differentiator. What separates machines after that is resilience, recovery speed, self-correction, consistency and cost of operation.當成功率已經高於 99%，它就不再是差異點。之後真正分開機器的是韌性、回復速度、自行修正能力、一致性與運作代價。
THREE LEVELS
What each level is, and what it is not三個層級各自是什麼，又不是什麼
Each level forms a benchmark on its own. A Level 1 result says nothing about whether a task will succeed, and a Level 3 result says nothing about why it failed. The value comes from holding the mapping between them.每一層都能獨立形成 benchmark。Level 1 的結果說不出任務會不會成功，Level 3 的結果說不出為什麼失敗。價值來自維持它們之間的映射。
HOW THE LEVELS CONNECT
A worked decomposition一個實際的拆解
A long-horizon task breaks into atomic loops, each atomic loop into reusable actions, and each action into a reflex that can be triggered and measured on a bench. The chain is what lets a behaviour-level failure end up as a component specification.長流程任務拆成原子迴路，每個原子迴路拆成可重用動作，每個動作再拆成可以在治具上觸發並量測的反射。有了這條鏈，行為層的失敗才能落到組件規格上。
BOTTLENECK LOCALISATION
Seven segments, one closed loop七個環節，一條閉迴路
A failure at the end of a task rarely tells you where the system is limited. We walk the loop segment by segment, with a characteristic symptom and a diagnostic for each, until the limiting segment is isolated.任務最後失敗，通常說不出系統受限在哪裡。我們沿著迴路逐段檢視，每一段都有典型症狀與對應的診斷方法，直到把受限的那一段分離出來。
1
Observation觀測
Cannot see it, sees it late, occlusion and drift看不到、看太晚、遮蔽與漂移
2
Representation and memory表徵與記憶
Loses the goal, cannot hold self and world state忘記目標、無法保持自身與世界狀態
3
Policy and decision策略與決策
Hesitates, oscillates, over-sensitive to small change猶豫、震盪、對微小變化過度敏感
4
Actuator command致動命令
Poor discretisation, saturation, bad mode switching動作離散化不佳、飽和、模式切換不良
5
Dynamics動力學
Commanded force does not become the intended motion指令的力量沒有變成預期的動作
6
Environment環境
Objects move, contact events, the target relocates物體狀態變化、接觸事件、目標移動
7
Next observation下一觀測
The new observation returns to stage one新的觀測回到第一段
WHAT COMES OUT
A capability fingerprint, not a score一份能力指紋，而非一個分數
A single number collapses everything that matters into one axis, and machines that behave very differently end up next to each other. We report the shape instead: where a machine holds, where it gives way first, and how far the gap opens as conditions move.單一數字會把所有重要的事壓成一個軸，行為差很多的機器會被排在一起。我們回報的是形狀：這台機器在哪裡撐得住、在哪裡先讓步，以及條件變動時差距拉開多少。
The axes shown here are illustrative. The real set is defined per application profile.此處的軸線為示意。實際軸線依每一份 Application Profile 定義。
All good都很好
Low discriminating power. This item becomes a basic conformity check and stops consuming effort.鑑別率低。這一項轉為基本符合性測試，不再投入資源。
All poor都很差
A shared bottleneck. This is usually where an architecture- or specification-level opening sits.共同瓶頸。這裡通常藏著架構或規格層級的機會。
Widely different差異很大
The most valuable outcome. Adjust conditions further, open the gap, and drive to the capability boundary.最有價值。繼續調整條件、拉開差距，追到能力邊界。
Test contracts, metric definitions, fixtures and reference sets sit outside this public brief.測試合約、指標定義、治具與參考集不在本公開概要內。
RESTRICTED
Held back until a briefing在說明之後才提供
The material below exists and is current. It stays off the public site while the company is in stealth, and is walked through directly rather than sent as a file.以下資料都已完成且持續更新。公司在 stealth 階段期間不放在公開網站上，並且以當面說明的方式進行，而非寄送檔案。
Platform architecture and operating loop平台架構與營運循環
Baseline, single-bottleneck hypothesis, paid retest, before/after evidence, and the data flywheel behind them.Baseline、單一瓶頸假說、付費複測、Before／After 證據，以及背後的資料飛輪。
Available in briefing說明會中提供
Benchmark map and differentiation metricsBenchmark 地圖與鑑別指標
Where a test sits on the system-integration and adoption-readiness axes, and what gets compared once success rate is above 99%.測試在系統整合與應用採用兩軸上的位置，以及成功率超過 99% 之後真正比較的項目。
Available in briefing說明會中提供
From benchmark to productisation從 Benchmark 到產品化
The eleven-stage path from SOTA calibration to a reference architecture and a restricted-domain product gate.從 SOTA 校準到參考架構與受限場景產品化的十一段流程。
Available in briefing說明會中提供
Positioning against adjacent approaches與相鄰做法的定位比較
Where buyer-driven differentiation testing sits relative to standards bodies, competitions and commercial trials.買方驅動的鑑別測試，相對於標準機構、競賽與商業實測的位置。
Available in briefing說明會中提供
First platform: autonomous flight第一個平台：自主飛行
Why the first differentiation platform is built on unmanned aircraft, and the minimum viable test package.第一個鑑別平台建立在無人機上的理由，以及最小可行的測試組合。
Available in briefing說明會中提供
Benchmark pipeline and scenario generation測試流程與場景生成
Come and measure the things nobody has measured yet來量那些還沒有人量過的東西
Most of this work does not exist as a job description anywhere else. We build the fixtures, the models and the machines that make robot capability comparable — and then we go and find where it breaks.這裡大部分的工作，在別的地方找不到對應的職務說明。我們把治具、模型與機器做出來，讓機器人的能力可以被比較，然後去找出它在哪裡壞掉。
01
Physical AIPhysical AI
Contact-rich manipulation, force and position control, and getting a learned policy to hold up on real hardware.接觸密集的操作、力與位置控制，以及讓學習出來的策略在真實硬體上站得住。
02
World ModelWorld Model
State representation, prediction under partial observation, and keeping a causal chain intact across a long task.狀態表徵、部分觀測下的預測，以及讓長流程任務的因果鏈不斷掉。
03
Autonomous Systems自主系統
Planning and decision under uncertainty, safe sets, failure detection, and recovery that is itself assessed.不確定條件下的規劃與決策、安全集合、失效偵測，以及本身也要被評估的回復動作。
04
Robotic Systems機器人系統
Whole-machine integration, sensing, real-time control, and the test infrastructure that proves any of it.整機整合、感測、即時控制，以及能證明這一切的測試基礎設施。
05
Electromechanical design機電系統設計
Mechanisms, actuation, fixtures and measurement rigs — the hardware that makes a number repeatable.機構、致動、治具與量測平台——讓一個數字可以重複的那些硬體。
EXPERIENCE實務經驗
Three years or more, hands on三年以上，實際動手做過
EDUCATION學歷
Bachelor's degree or above, science or engineering理工學士以上
A CV, and one paragraph on something you built or measured yourself — what the number was, what conditions it held under, and where it stopped holding.履歷，加上一段文字：講一件你自己做過或量過的東西——當時的數字是多少、在什麼條件下成立、以及在哪裡開始不成立。
A shortcut一個捷徑
If any of the sixteen bottlenecks on this site is a problem you have personally fought, quote its number in the first line. That tells us more than a job title does.如果本站十六項瓶頸裡，有哪一項是你親自纏鬥過的問題，第一行就寫下它的編號。那比任何職稱都更能說明你的能力。
We reply to everyone. If the fit is not there now we will say so, and we keep good CVs on file with your permission.我們每一封都會回覆。目前不合適的話我們會直說，並在徵得同意後保留履歷。
Ed Wei · RobotSys.ai · Taiwan Robotic System Co. · TaiwanEd Wei · RobotSys.ai · 台灣自主機器人系統公司 · 台灣
Investors投資人
Please make contact first. We arrange a briefing rather than circulate a deck, and the restricted material is walked through in that session.請先來信聯絡。我們安排說明，而非流通簡報檔，受限資料會在該場說明中逐項說明。
Helpful in a first message: fund or firm, stage focus, and whether robotics or semiconductor is your core area.第一封信中如果能包含：機構名稱、關注階段，以及您的核心領域偏機器人或半導體，會很有幫助。
Robot makers, buyers and research groups機器人廠商、買方與研究單位
If you have an application whose acceptance conditions are hard to write down, that is exactly the conversation we want. Tell us the task and what failure costs you.如果您有一個驗收條件很難寫清楚的應用，那正是我們想談的。請告訴我們任務內容，以及失敗的代價。
We are also open to joint definition of an application profile before any testing begins.在任何測試開始之前，我們也歡迎共同定義 Application Profile。
All figures on this site describe industry-level challenges and are shared for discussion.本站所有圖示均為產業層級的挑戰說明，供討論使用。
