The handoff nobody has timed

In April 2026 thirteen authors published a system that noticed one of its robots was missing.

ABot-Claw — arXiv 2604.10096, v2 of 17 April 2026 — runs on a Unitree G1 humanoid, a Unitree Go2 quadruped and an Agilex Piper arm. In §4.3.2 a user asks for the quadruped's status. The system tries to reach it; the service is unavailable. Verbatim: "Instead of terminating, ABot-Claw queries the shared memory to recover the quadruped's last known location and uses that information to reassign the task." The humanoid is dispatched to that location to inspect it. (The paper's byline reads AMAP CV Lab; the arXiv record lists thirteen authors and no affiliation fields, and this article goes no further than the byline.)

That is a real system doing a real thing on real hardware, and the closest published case to a capability we sell. The quadruped never started the task. It is the subject of the inspection, not the sender. Nothing in flight was handed anywhere: a task that could not be assigned to one body was assigned to another, and a shared memory supplied the coordinate. Memory-mediated cross-embodiment reassignment, not a handoff.

We went looking for a number attached to it — attempts, successes, duration — and the absence has a precise shape rather than a broad one. Both halves were counted on machines this year. DUET passed an object between two genuinely different bodies on real hardware with a denominator attached: a successful handover in 3 of 10 trials. RoboHarness measured handoff through a shared memory across 135 real-robot experiments, between two policies inside one body. The sentence to defend is this one. The handoff is measurable — two teams measured the halves of it in 2026 — and nobody has measured the whole of it on a machine. That includes us.

What the evaluation sections say

Three systems sit nearest the capability. We read each one's evaluation section in full on 11 September 2026, from the arXiv full text rather than an abstract, because a negative established from abstracts is not established.

ABot-Claw contains no table anywhere in the document [primary, §4 read in full — negative]. §4 is five demonstrations, each a narrated single run, with no trial count, success rate, duration, latency, baseline, ablation or benchmark for any of them. Nothing in the paper was measured, so the handoff is not a special case. It does describe the capability we came for. §2.2, the method section, offers it as an illustration of the task-allocation strategy: "For instance, in an object transfer task, a mobile robot may deliver an object to a workstation, after which the runtime hands over the object pose and execution context to a fixed manipulator…" Note the modal verb. Architecture, not a run, and not one of the five demonstrations.

HoloAgent-0 (arXiv 2606.23565v1, 22 June 2026) deploys on a Unitree G1, an R1 humanoid and a wheeled dual-arm mobile manipulator. Its entire cross-robot coordination result is one figure caption: "(c) Cross-Robot Coordination: route one robot while another performs a dance skill." Two robots doing two different things at once is parallel dispatch through shared task state, not a handoff. The paper does publish numbers — success rate and SPL for long-horizon navigation, accuracy for its 3D semantic map — and none attach to cross-robot coordination. It separates the two itself, and says why [primary, §5.1 + §5.4 — negative]:

"For full-stack embodied-agent behavior, we report qualitative execution traces rather than a single end-to-end success benchmark, because manipulation, whole-body motion, and cross-embodiment coordination use heterogeneous hardware and are not yet standardized under one repeatable protocol."

§5.4 adds that a rigorous end-to-end benchmark for cross-embodiment collaboration "remains future work". The credit for that belongs to its authors.

RoboOS-NeXT is arXiv 2510.26536, dated 30 October 2025, not 2026. Its §IV-G real-hardware work is three qualitative scenarios with no counts, and its predecessor RoboOS (arXiv 2505.03673, May 2025) carried the same three — same scenarios, same robots, no counts — five months before it [primary, §IV-G and RoboOS §4.3 — negative]. It does measure the robot-offline case, which is precisely ABot-Claw's: Table III, household, 87.6 percent success against a memory-less baseline of 44.5 percent — in a mock setting that, in its own authors' words, "abstracts away physical uncertainties and focuses on system effectiveness". The qualifier travels with the number every time, because the physical uncertainties are where a handoff breaks.

Five rungs

Rung one: cross-embodiment policy transfer, measured at a scale nothing else here approaches, offline. Open X-Embodiment, arXiv 2310.08864: "We assemble a dataset from 22 different robots collected through a collaboration between 21 institutions, demonstrating 527 skills (160266 tasks)." That is a dataset and a policy checkpoint — one model trained offline on pooled data, then evaluated on several robots. Nothing is handed from one robot to another, and nothing happens during a shift. Policy transfer and task handoff share a prefix and nothing else.

Rung two: knowledge transferred between two robots, measured, and it needs a restart. A Semantic Autonomy Framework for VLM-Integrated Indoor Mobile Robots (arXiv 2605.02525, 4 May 2026) reports 100 percent semantic transfer accuracy from one robot to another: 33 of 33, 95 percent CI [0.894, 1.000], across 82 scenario-level decisions in three sessions. The denominator is published. Both robots are differential-drive units from one platform family, deliberately non-identical below the stack — "Their low-level hardware differs in compute distribution, drivetrain electronics, encoder type, and IMU availability." The transfer is not in flight. Step seven of the paper's own nine-step mechanism is "Robot B loads the same digest at startup", and §5.6 records that "Memory updates are performed manually between experimental sessions by running refresh_memory.sh." The best-measured cross-robot memory transfer in the literature is an offline batch job and a reboot.

Rungs three and four are the two halves, below. Rung five is their intersection, where the published numbers stop.

Rung 1 — Policy transfer

• Open X-Embodiment (arXiv 2310.08864): 22 robots, 21 institutions, 527 skills, 160,266 tasksMeasured offline. Nothing is handed from one robot to another, and nothing happens during a shift.

Rung 2 — Knowledge transfer, same morphology

• Semantic Autonomy Framework (arXiv 2605.02525): 33 of 33 semantic transfers, 95% CI [0.894, 1.000]Two differential-drive robots from one platform family. Robot B loads the digest at startup and memory updates are run by hand between sessions: a batch job and a reboot, not a transfer in flight.

Rung 3 — Physical handover, two bodies

• DUET (arXiv 2606.20990): successful handover in 3 of 10 trials on real hardwareInitial grasp 8 of 10; robot-only baselines 2 of 10 and 1 of 10. A Unitree G1 and a Dexmate Vega1, driven by one centralised policy: no shared memory, no task state.

Rung 4 — Memory-mediated handoff, one body

• RoboHarness (arXiv 2607.18060): 135 real-robot experiments, 86.7% on the Bridge structure task80.0% after a partially built structure is dismantled, 73.3% under 5–10% pose-estimate error, 66.7% when required blocks are re-hidden mid-execution. Per-condition denominators on a machine — but heterogeneous policies on one arm, not two bodies.

Rung 5 — In-flight handoff, two bodies, shared memory

• Demonstrated on hardware, uncounted — ABot-Claw (arXiv 2604.10096)Five demonstrations on three real platforms, each a narrated single run. No table anywhere in the document: no trial count, success rate, duration, latency, baseline or ablation. And the case narrated is reassignment, not a handoff — the second robot never started the task.
• Measured in a mock setting — RoboOS-NeXT (arXiv 2510.26536): 87.6% under robot-offline, memory-less baseline 44.5%Mock setting: in the authors' own words it “abstracts away physical uncertainties and focuses on system effectiveness”, and the physical uncertainties are where a handoff breaks. Its three real-hardware scenarios carry no counts, and its predecessor carried the same three — same scenarios, same robots, no counts — five months before it.
• Measured on real hardware — nothingNo trial count, no success rate and no duration on any machine, anywhere we could reach, for an in-flight task handed between two different bodies through a shared memory. Ours included: no Tidewell robot has handed anything to another Tidewell robot.

• A published count from real robots — a trial denominator on rungs 2 to 4, and on rung 1 the size of a dataset assembled from them
• Not a measurement on real hardware: a mock setting, a demonstration with no count, or nothing at all

The two halves, and the word for the gap

DUET (arXiv 2606.20990v1, 18 June 2026) runs a Unitree G1 humanoid and a Dexmate Vega1 mobile manipulator — two genuinely different bodies. Its fourth task: "T4: Doll Passing focuses on coordinated handovers and grasp transfer. Vega1 secures and transfers a doll directly to G1, requiring synchronized release and capture maneuvers." Its evaluation: "To evaluate policy performance, we execute 10 independent physical hardware trials per task." Table 3 scores each trial against two milestones, the second literally named Successful Handover. Initial grasp, 8 of 10. Successful handover, 3 of 10, against robot-only baselines of 2 of 10 and 1 of 10.

Three of ten is a low number worth more than a high one with nothing behind it — the argument of a success rate without a denominator, not this article's to repeat. What DUET is not is the thing on our product page: no shared memory, no task state. One centralised policy predicts actions for both bodies at 10 Hz from both robots' cameras and joint positions; on asynchronous delay the system freezes both. Two bodies, one controller, one physical pass.

RoboHarness (arXiv 2607.18060v2, 28 July 2026) collides with our vocabulary exactly: "inter-policy handoff", "stable policy handoffs", and a component called the Memory Bridge, which "retrieves execution trajectories associated with the next policy, estimates its in-distribution state region, and guides the robot toward that region without joint policy retraining". And it counted: "We conducted 135 real-robot experiments: 15 for each of five target structures and 15 under each of four disturbance settings." On the Bridge structure task — one of the five it builds, not the Memory Bridge — success is 86.7 percent; 80.0 percent after a partially completed structure is dismantled; 73.3 percent under 5 to 10 percent pose-estimate error; 66.7 percent when required blocks are re-hidden mid-execution. Per-condition denominators, injected disturbances, on a machine.

"Heterogeneous" there means heterogeneous policies — a π0.5 VLA and a TAMP planner — on one arm. Not cross-embodiment. But proof that this class of handoff is measurable under disturbance, with a denominator attached.

Put the two side by side. The same research community measured the handoff carefully where one body owns both ends of it, and measured a physical pass carefully where two bodies share an object and no memory. Where those conditions meet — an in-flight task handed between two different bodies through a memory both of them read and write — there is no trial count, no success rate and no duration on real hardware that we could reach. Ourselves included: we have not run it either, and nothing in this article is a Tidewell measurement.

Somebody has defined the metric, in simulation

Our own research notes said that nobody had defined a handoff benchmark. That was already false on the day we wrote it, 11 September 2026: the definition had been on arXiv for two weeks and we had not found it.

CoCoBench (arXiv 2608.28266, 28 August 2026) is 897 oracle-validated instances in AI2-THOR across four coordination constructs, evaluated on 11 MLLMs. The fourth construct is handoff, with 224 instances, defined explicitly: "D4: Handoff coordination covers transport tasks where agents pass objects through an intermediate surface, such as a table or container… Let h be the number of successful buffer handoffs, o be the number of overflow events, and w be the number of empty-buffer waiting events." The score is C_D4 = 1 − min(1, (o + w) / h). Its motivating sentence is our own argument in someone else's words, published before ours: overall success rates "can obscure coordination failures such as duplicated work, violations of ordering constraints, resource contention, and desynchronized handoffs."

What it does not cover matters as much. There is no robot hardware; the agents are MLLM planners over a fixed high-level skill interface; and h, o and w count buffer events, so there is no clock, no body and no memory read [primary — simulation, AI2-THOR].

An independent survey arrives at the same place from the other direction. When Multi-Robot Systems Meet Agentic AI (arXiv 2606.27929, 26 June 2026), §III-C: "The limitation is that this sharing is usually offline and model-centric… Learning-era sharing therefore made robot experience scalable, but not yet a real-time coordination mechanism." Its own case study is a fourth near-miss — memory inheritance in simulation, 166 target-object navigation tasks, four robots that "differ in their history rather than their body or policy", 24.1 percent success with no memory against 77.1 percent for the robot inheriting the merged team memory [primary — simulation]. The inheritance happens on arrival, not in flight.

What a measurement would have to contain

Here is the protocol, written so somebody else can run it before we do, and built against work that already exists: CoCoBench's h, o and w; RoboHarness's per-condition denominators; and the task-state ledger the survey above proposes in §IV-B, whose scheduler "triggers reassignment when a robot goes offline or when work has already been completed by another member."

Denominator. Attempted handoffs. An attempt begins the moment the sending body's task state is marked for transfer, and is counted whether or not a receiver exists. Counting only the attempts that found a receiver is how this number gets quietly inflated.

Success. The receiving body completes the original task's remaining acceptance criteria without re-perceiving what the sender had already written to memory. Completion after re-perception is a partial, because a receiver that re-perceives everything has not used the memory and would have done as well with none.

Clock. Starts at the sender's write of the transfer record. Two stops, not one: memory read complete, and first actuated motion against the transferred goal. A slow read and a slow approach are different defects, and a single duration hides which one you have.

Failures, named in advance. Memory read failure, pose mis-registration, capability mismatch, dropped object, timeout, silent abandonment. The last will not appear in any table unless it is in the taxonomy before the trials start.

Interventions. Counted separately, never netted out of the success rate. Every figure travels with its trial count. [inference — this protocol is ours, and nothing in it has been run]

The constraint is agreement on what to count, not instrumentation. One real-robot leaderboard reported this year ran 1,088 runs over 25,627 episodes [secondary], every one of them a single robot doing a single task. The field can run twenty-five thousand episodes once it has decided what a trial is.

One attempted cross-embodiment handoff — the unit to be counted

1. 1. TriggerTask state marked for transfer. The clock starts here, and so does the denominator.
2. 2. Memory read
   Clock stop 1Memory read complete
3. 3. Capability matchWhether the receiving body can finish what remains.
4. 4. TransferThe receiver takes the task on.
   Clock stop 2First actuated motion against the transferred goal
5. 5. VerificationSuccess is the receiving body completing the original task's remaining acceptance criteria without re-perceiving what the sender had already written to memory. Completion after re-perception is a partial.

The clock starts at the sender's write of the transfer record. Two stops, not one: a slow read and a slow approach are different defects.

Failures, named in advance

• Memory read failure
• Pose mis-registration
• Capability mismatch
• Dropped object
• Timeout
• Silent abandonmentThe one that will not appear in any table unless it is in the taxonomy before the trials start.

Proposed protocol — nothing measured. The denominator is attempted handoffs: an attempt begins the moment the sending body's task state is marked for transfer, counted whether or not a receiver exists.

• A stage of the proposed sequence
• A clock stop the protocol requires — both of them, because one duration hides which defect you have
• A failure mode named before the trials start. The list is flat on purpose: nothing has been run, so the protocol does not yet say which stage each one belongs to

What was searched, and what was not

The zero above is bounded, and the bound travels with it. On 11 September 2026 we ran about twenty queries against the arXiv API's full-record metadata search. Exact-phrase searches for "cross-embodiment handoff", "task handoff" and "handoff success rate" return 0 records across all of arXiv; "handoff latency" returns ten, and not one of them is a robot-to-robot task handoff — eight are wireless-networking papers, one is an on-device LLM KV-cache paper, and one is a control paper about computation and handoff latency inside a single sampled-data controller. We walked the Semantic Scholar citation graph for RoboOS-NeXT, and ran four web searches in English and three in 中文 [primary — negative, API query]. Those exact-phrase queries were re-run on 12 September 2026 and returned the same counts.

Not searched: IEEE Xplore and the ACM Digital Library, both paywalled to automated fetch. One record surfaced through the citation graph and is named here so the negative is honest — Helten and Wolf, IEEE UR 2026, leader–follower knowledge sharing between robots with heterogeneous sensing, evaluated by dataset playback rather than on hardware according to its abstract, full text closed and unread by us [secondary — abstract only]. Also not searched: ICRA 2026 and IROS 2026 proceedings not mirrored on arXiv, vendor test reports, Japanese-language work, patents and theses.

The three systems' own project pages and repositories were the likeliest place to find a count the PDFs do not carry, so we fetched all of them on 12 September 2026. None carries a trial count, a success rate or a duration for anything cross-embodiment. ABot-Claw's repository is a deployment guide with no results section. RoboOS's project site reproduces the abstract and no numbers. HoloAgent-0's page does publish numbers the paper also publishes — simulated HM3D-ObjNav success rate and SPL, and real-robot goal-reaching at one metre — all of them navigation, and its cross-robot coordination entry is the same one-line video caption as Figure 2(c), with nothing attached [primary — pages read in full, 12 September 2026].

What would falsify this: one published trial count, success rate or duration for an in-flight task handed between two different bodies through a memory both of them read and write, on real hardware. One document is enough. If it exists in a venue named in the paragraph above, we have not seen it, and the piece that replaces this one begins by saying we were wrong.

A fourth demonstration sits outside the literature. NetEase trade coverage of WRC 2026, 20 August 2026, describes 越疆 Dobot's 一脑多体 ("one brain, many bodies") platform on a production line: "人形机器人、四足机器狗与协作臂无缝接力，完成从分拣、转运到装配的完整工序" — a humanoid, a quadruped robot dog and a collaborative arm relay seamlessly, from sorting through transport to assembly. The same article describes 里工 validating "跨机器人工序交接", cross-robot process handover, between two 里掂 F1 units of the same morphology. No trial count, success rate or duration is given for either, and the Dobot claim could not be corroborated on the vendor's own channel in this pass [secondary, single source]. That bound is two Chinese-language searches and one trade article read in the original, which makes it a weaker negative than the arXiv one.

Our own page, corrected twice

Tidewell Brain lists cross-embodiment handoff in its task list as a design target rather than a demonstrated capability. It reached that form in two corrections, both ours and both dated.

The first reclassified the handoff from something we do to something we are building. The second, made on 11 September 2026 out of the research behind this article, fixed what the first had left standing. The page said three systems demonstrated the handoff on real hardware during 2026. Wrong on the date: RoboOS-NeXT is October 2025, and its predecessor carried the same scenarios in May 2025. Wrong on the count: HoloAgent-0's case is parallel dispatch. And written before we had read DUET or RoboHarness, which is why it claimed an absence broader than the evidence supports. The same sentence on Crew now carries the corrected version. Our notes on what breaks in a shared memory and on the empty square are unchanged.

For anyone watching a two-robot video and trying to tell the capability from the edit, the ladder is the question set. Was the second robot driven by a controller already driving the first — rung three? Did it read the task state from a memory the first one wrote — rung five? Did it read that memory while the task was running, or at startup — rung two? Then ask for the denominator. All four are answerable from a method note two sentences long.

Our own position, plainly: W1 and B1 are prototypes, C1 is in development, Crew is not running on any site, and no Tidewell robot has handed anything to another Tidewell robot. Publishing the protocol above has a cost we can see — a competitor can run those definitions before we do, and we will be held to them when our own numbers arrive. When we have run it, we will publish the trial count beside the result, and the failures in the same table.