robo-si / M2

M2 · 长物协作搬运Cooperative long-object transport

多台机器人从地面合作抬起一只 2.25 × 0.35 × 0.24 m 的长箱，穿过通道运到地面目标区或接收架，稳定交付、撤手并停稳。六项配置覆盖双 G1、G1 + Stretch 3 和三台 G1 三种团队，检验双端抓持、同步起放、转弯运输、异构分工、加重适应以及清障与搬运的先后顺序。所有移动须经关节、轮子与真实接触产生，不得预抓持、运行中重设位姿、隐藏承托或吸附货物；完整成功必须来自一次连续运行。Several robots cooperatively lift a 2.25 × 0.35 × 0.24 m long box from the floor, carry it through a corridor to a floor goal zone or a receiving rack, deliver it stably, let go and come to rest. The six configurations use three teams (two G1s, G1 + Stretch 3, three G1s) and test two-ended grasping, synchronized lifting and lowering, turning while carrying, heterogeneous role allocation, adaptation to a heavier load, and the order of obstacle clearing versus transport. All motion must come from joints, wheels and real contact with no pre-grasp, runtime pose resets, hidden support or attachment, and a full success must come from one continuous run.

验收范围Accepted scope

作者于 2026-11-02 确认 M2 本轮参考交付完成，覆盖全部 6 项配置：每项都有一次连续完整的物理成功见证及同源码复测（M2-1～M2-5 各 3/3，M2-6 全组 3/6）。正式人工视频/判据审阅与第二人独立复现仍未签署，也未新增正式 benchmark 准入；专家使用仿真状态反馈，结果限于固定实例，不证明随机场景、跨环境稳定性、实时运行或硬件部署。原异构 L 型路线、拖板与工具发现不在本轮范围，历史失败全部保留。On 2026-11-02 the author confirmed the current M2 reference delivery as complete for all six configurations: each has one continuous full physical success witness plus same-source repeats (3/3 for each of M2-1 to M2-5, 3/6 for the whole M2-6 group). Formal human video/criteria review and second-person independent reproduction are still unsigned and no formal benchmark admission is claimed; the expert uses simulator state feedback and the results cover fixed instances only, not randomized scenes, cross-environment stability, real-time operation or hardware deployment. The original heterogeneous L-shaped route, the dolly and tool discovery are out of scope, and all historical failures are retained. (2026-11-02)

变化的条件轴Condition axes

只变化三类条件：货物质量（0.25 或 0.5 kg，为仿真实例参数，不是硬件额定负载）、团队（双 G1、G1 + Stretch 3、三台 G1）和场景布局（L 型通道地面交付、直通道接收架交付、L 型通道加可移动障碍）。有效对照为 M2-1↔3 质量、M2-2↔4 异构场景质量、M2-3↔5 障碍、M2-5↔6 团队规模；双 G1 与异构任务场景不同，只能作策略描述比较，不作 embodiment 因果结论。Only three things vary: cargo mass (0.25 or 0.5 kg, a simulation instance parameter rather than a hardware load rating), the team (two G1s, G1 + Stretch 3, three G1s) and the layout (L-shaped corridor with floor delivery, straight corridor with receiving-rack delivery, L-shaped corridor with a movable obstacle). The valid comparisons are M2-1 vs 3 (mass), M2-2 vs 4 (mass in the heterogeneous scene), M2-3 vs 5 (obstacle) and M2-5 vs 6 (team size); the dual-G1 and heterogeneous tasks use different scenes, so they support only descriptive strategy comparison, not causal claims about embodiment.

Embodiments

两台 Unitree G1 人形（冻结模型，Dex3 三指手），自由行走Two Unitree G1 humanoids (frozen model, Dex3 three-finger hands), walking freely
一台 G1（Dex3 手）+ Stretch 3（SE3 / DW3 / SG3）：差速轮底盘、升降与伸缩臂，以原生夹爪指端承托；Stretch 4 不算已验证替代One G1 (Dex3 hands) + a Stretch 3 (SE3 / DW3 / SG3): differential-drive base, lift and telescoping arm, supporting the box with its native gripper fingertips; Stretch 4 is not counted as a validated substitute
三台相同版本的 Unitree G1 人形（Dex3 三指手）Three Unitree G1 humanoids of the same version (Dex3 three-finger hands)

统计Counts

6 个已验收任务／配置，其中 0 个拒答；6 个有倍速视频。6 accepted tasks/configurations, 0 of them refusal cases; 6 with sped-up videos.

证据备份Evidence backup

本族 6/6 个已验收配置的参考证据已归档：21,739 个文件，19.4 GiB。Reference evidence for 6/6 accepted configurations is archived: 21,739 files, 19.4 GiB. 另归档非参考开发运行 108,077 个文件（42.0 GiB）。Non-reference development runs are archived as well: 108,077 files (42.0 GiB). 2026-11-03 对远端逐文件核对大小与 SHA-256。Every remote file's size and SHA-256 were checked on 2026-11-03.

M2-1 M2-2 M2-3 M2-4 M2-5 M2-6

8×

8 倍速 · 原始 175 s，播放 22 s8x speed · original 175 s, plays in 22 s

M2-1双 G1 轻载通过 L 型通道Two G1s carry the light box through the L-shaped corridor

Unitree G1执行任务Solve the task

条件Conditions: 0.25 kg 长箱（2.25 × 0.35 × 0.24 m）从地面开始，无预抓持；冻结 L 型通道净宽 1.8 m、墙高 1.5 m，两台 G1 位于箱两端的原起始站位。模型、接触参数、目标和历史 L1 评分版本沿用已发布的 M2 第一层参考包；固定实例（同场景同种子），专家使用仿真状态反馈。0.25 kg long box (2.25 × 0.35 × 0.24 m) starting on the floor with no pre-grasp; frozen L-shaped corridor 1.8 m wide with 1.5 m walls, the two G1s at their original start poses at the two ends of the box. Model, contact parameters, goal and the historical L1 scoring version come from the published M2 layer-1 reference package; fixed instance (same scene and seed), and the expert uses simulator state feedback.
任务设计Task design: 两台 G1 要把地面长箱运过转角，完整放入地面目标区，稳定交付、撤手并停稳。专家计划为：观察检查 → 两端双手抓持 → 同步起箱 → 在转角协调身体朝向、步态与长箱端点并运输 → 同步下放 → 释放停车；抓持以实际双手接触与货物承重确认。The two G1s must carry the floor box around the corner, place it wholly inside the floor goal zone, release it and come to rest. The expert plan is: inspect → two-handed grasp at both ends → synchronized lift → carry around the corner while coordinating body heading, gait and the box ends → synchronized lowering → release and stop; the grasp is confirmed by real hand contact and load support.
工具Tools: 无None
Embodiment: 两台 Unitree G1 人形（冻结模型，Dex3 三指手），自由行走Two Unitree G1 humanoids (frozen model, Dex3 three-finger hands), walking freely
是否拒答Refusal: 否No
见证结果Witness result: 成功：选定运行一次连续完成起箱、转弯运输、地面交付与释放，仿真用时 174.36 s；同源码同种子三次复测 3/3 通过，用时一致。仅证明固定实例可执行，不代表随机覆盖或跨环境稳定；人工签审与第二人复现未完成。Success: the selected run lifted the box, carried it around the corner, delivered it on the floor and released it in one continuous run, 174.36 s of simulated time; three same-source, same-seed repeats passed 3/3 with identical timing. This shows the fixed instance is executable, not random coverage or cross-environment stability; human sign-off and second-person reproduction are pending.

8×

8 倍速 · 原始 267 s，播放 33 s8x speed · original 267 s, plays in 33 s

M2-2G1 + Stretch 3 轻载接收架交接G1 + Stretch 3 deliver the light box onto a receiving rack

Unitree G1Stretch 3执行任务Solve the task

条件Conditions: 0.25 kg 长箱从地面开始，无预抓持；采用作者授权的接收架 v3 场景：直通道净宽 2.35 m、墙高 1.5 m，两个固定可碰撞接收支点各 0.14 × 0.65 m、顶面高 0.64 m、间距 1.6 m，参考搬运距离 2.4 m。机器人不得借接收架支撑身体；固定实例，专家使用仿真状态反馈。0.25 kg long box starting on the floor with no pre-grasp, in the author-authorized receiving-rack v3 scene: straight corridor 2.35 m wide with 1.5 m walls and two fixed, collidable receiving supports, each 0.14 × 0.65 m with a 0.64 m top, 1.6 m apart; reference carry distance 2.4 m. Robots may not lean on the rack; fixed instance, and the expert uses simulator state feedback.
任务设计Task design: G1 双手握近端并控制箱体姿态，Stretch 从侧方以夹爪指端插到箱底承托，靠差速轮、升降和伸臂配合移动。专家先让 G1 单端起箱、Stretch 按实测净空插托，调平并站起；完成步态交接、起步后再平滑抬高 60 mm 沿直通道搬运，到架上方下放，按实测载荷卸下叉托，确认两个支点均受载才收臂，G1 张手撤腕，双机与货物脱离后停稳。The G1 holds the near end with both hands and controls the box attitude, while the Stretch slides its gripper fingertips under the box from the side and moves with its wheeled base, lift and arm. The expert has the G1 lift one end and the Stretch insert its support at the measured clearance, then levels and stands; after the gait handoff and start it raises the carry height 60 mm smoothly, carries along the straight corridor, lowers onto the rack, unloads the fork by measured load, retracts only once both supports carry load, and the G1 opens its hands so both robots disengage and stand still.
工具Tools: 接收架：两个固定、可碰撞的接收支点（公开环境设施，不是隐藏承托或运输工具）Receiving rack: two fixed, collidable supports (a public environment fixture, not hidden support or a transport tool)
Embodiment: 一台 G1（Dex3 手）+ Stretch 3（SE3 / DW3 / SG3）：差速轮底盘、升降与伸缩臂，以原生夹爪指端承托；Stretch 4 不算已验证替代One G1 (Dex3 hands) + a Stretch 3 (SE3 / DW3 / SG3): differential-drive base, lift and telescoping arm, supporting the box with its native gripper fingertips; Stretch 4 is not counted as a validated substitute
是否拒答Refusal: 否No
见证结果Witness result: 成功：选定运行一次连续完成地面起箱、直线搬运、架上卸载与双机脱离，仿真用时 266.62 s；同一冻结源码三次运行 3/3 通过（266.62 / 265.02 / 274.00 s）。复测是在观察到首个见证后追加的，不是预先声明的成功率；旧的提前抬高控制组为 1/3，原 L 型路线与 v2 接收架的失败均保留，不追认成功。Success: the selected run picked the box off the floor, carried it straight, unloaded it onto the rack and both robots disengaged in one continuous run, 266.62 s of simulated time; three runs of the same frozen source passed 3/3 (266.62 / 265.02 / 274.00 s). The repeats were added after the first witness was observed and are not a predeclared success rate; the earlier lift-before-handoff control group stands at 1/3, and failures on the original L route and the v2 rack are retained and not counted as successes.

8×

8 倍速 · 原始 176 s，播放 22 s8x speed · original 176 s, plays in 22 s

M2-3双 G1 加重适应Two G1s with the heavier box

Unitree G1执行任务Solve the task

条件Conditions: 与 M2-1 相同的 L 型通道、团队与起始站位，唯一变化是长箱质量增至 0.5 kg（尺寸与重心定义不变）。此前 1 kg 的失败记录保留，不解释为硬件能力上限；固定实例，专家使用仿真状态反馈。Same L-shaped corridor, team and start poses as M2-1; the only change is the box mass, raised to 0.5 kg (same size and centre-of-mass definition). Earlier 1 kg failures are retained and not read as a hardware limit; fixed instance, and the expert uses simulator state feedback.
任务设计Task design: 沿用双端抓持、起箱、转弯运输、下放与释放的结构，但要在加重后重新验证抓持滑移、载荷分配、关节力矩和步行稳定性；阶段推进依据实际状态，不要求两台平均承重。It keeps the two-ended grasp, lift, turning carry, lowering and release structure but must re-verify grip slip, load sharing, joint torques and walking stability under the heavier load; phases advance on the actual state and the two robots need not share the load equally.
工具Tools: 无None
Embodiment: 两台 Unitree G1 人形（冻结模型，Dex3 三指手），自由行走Two Unitree G1 humanoids (frozen model, Dex3 three-finger hands), walking freely
是否拒答Refusal: 否No
见证结果Witness result: 成功：选定运行一次连续完成搬运、交付与释放，仿真用时 175.88 s；同源码同种子三次复测 3/3 通过（175.88 / 175.88 / 177.82 s）。仅为固定实例复测，不推出工具收益或未知物性辨识能力。Success: the selected run carried, delivered and released the box in one continuous run, 175.88 s of simulated time; three same-source, same-seed repeats passed 3/3 (175.88 / 175.88 / 177.82 s). Fixed-instance repeats only; no claim about tool benefit or identifying unknown object properties.

8×

8 倍速 · 原始 269 s，播放 34 s8x speed · original 269 s, plays in 34 s

M2-4G1 + Stretch 3 加重接收架交接G1 + Stretch 3 deliver the heavier box onto the receiving rack

Unitree G1Stretch 3执行任务Solve the task

条件Conditions: 几何、机器人、初始工作区、接收架和终态条件与 M2-2 完全相同，仅长箱质量由 0.25 kg 改为 0.5 kg。固定实例，专家使用仿真状态反馈；此任务不再声称验证拖板机制选择。Geometry, robots, start workspace, receiving rack and terminal conditions are identical to M2-2; only the box mass changes from 0.25 kg to 0.5 kg. Fixed instance with simulator state feedback for the expert; the task no longer claims to test choosing a dolly mechanism.
任务设计Task design: 沿用“起箱—插托—站起—搬运—架上交接—释放”的高层结构，控制器按负载适应。重点是 G1 抓持与手臂余量、Stretch 实际承载、行走时的侧向扰动，以及卸载时两支点载荷和箱体俯仰：叉托下降不等于架子已承担货物，单支点接触也不能触发撤手。Same high-level structure (lift, insert support, stand, carry, rack handoff, release), with the controller adapting to the load. The key points are the G1 grip and arm margin, the Stretch actual load, lateral disturbance while walking, and the load on both supports and box pitch during unloading: lowering the fork does not mean the rack has taken the load, and contact on one support alone must not trigger letting go.
工具Tools: 接收架：两个固定、可碰撞的接收支点（公开环境设施，不是隐藏承托或运输工具）Receiving rack: two fixed, collidable supports (a public environment fixture, not hidden support or a transport tool)
Embodiment: 一台 G1（Dex3 手）+ Stretch 3（SE3 / DW3 / SG3）：差速轮底盘、升降与伸缩臂，以原生夹爪指端承托；Stretch 4 不算已验证替代One G1 (Dex3 hands) + a Stretch 3 (SE3 / DW3 / SG3): differential-drive base, lift and telescoping arm, supporting the box with its native gripper fingertips; Stretch 4 is not counted as a validated substitute
是否拒答Refusal: 否No
见证结果Witness result: 成功：选定运行一次连续完成交付，仿真用时 268.42 s，两个支点均实际受载、双机与货物接触为零；同一冻结源码三次运行 3/3 通过（268.42 / 265.06 / 268.42 s），终态、执行器与接触判据均未放宽。复测为观察首个见证后追加，不证明随机或跨平台鲁棒性；早期 v2/v3 失败保留。Success: the selected run completed delivery in one continuous run, 268.42 s of simulated time, with both supports actually loaded and zero robot contact on the cargo; three runs of the same frozen source passed 3/3 (268.42 / 265.06 / 268.42 s) with no relaxation of the terminal, actuator or contact criteria. The repeats were added after the first witness and do not show randomized or cross-platform robustness; earlier v2/v3 failures are retained.

8×

8 倍速 · 原始 260 s，播放 32 s8x speed · original 260 s, plays in 32 s

M2-5双 G1 先清障再搬运Two G1s clear an obstacle, then carry

Unitree G1执行任务Solve the task

条件Conditions: 在 M2-3（0.5 kg）基础上，L 型通道转角处增加一只已声明的可移动障碍（0.35 × 0.35 × 0.45 m、0.6 kg）和地面暂存区（0.50 × 0.55 m）。清障允许声明的膝/踝接触，货物仍只允许手/腕接触；固定实例，专家使用仿真状态反馈。Built on M2-3 (0.5 kg): a declared movable obstacle (0.35 × 0.35 × 0.45 m, 0.6 kg) is added at the corner of the L-shaped corridor, together with a floor staging area (0.50 × 0.55 m). Clearing may use declared knee/ankle contact, while the cargo still allows only hand/wrist contact; fixed instance, and the expert uses simulator state feedback.
任务设计Task design: 一台 G1 先接近障碍，用膝/踝把它推入暂存区，释放并确认障碍完整包含且静止，再返回搬运站位；随后两台完成起箱、转弯运输、下放与释放。成功须同时满足货物交付、障碍静止留在暂存区、全部机器人停稳；只推开一点或只完成清障都不算成功，若另一合法策略能绕过障碍也应接受并报告。One G1 first approaches the obstacle, pushes it into the staging area with its knee/ankle, releases it and confirms it is fully inside and still, then returns to its carrying pose; the two G1s then lift, carry around the corner, lower and release. Success requires cargo delivery, the obstacle resting inside the staging area and all robots standing still; nudging it partway or finishing only the clearing is not success, and a legal strategy that bypasses the obstacle would be accepted and reported.
工具Tools: 可移动障碍（0.35 × 0.35 × 0.45 m，0.6 kg，不是固定墙）Movable obstacle (0.35 × 0.35 × 0.45 m, 0.6 kg; not a fixed wall)地面暂存区（0.50 × 0.55 m），清障后障碍须稳定留在其中Floor staging area (0.50 × 0.55 m) where the cleared obstacle must come to rest
Embodiment: 两台 Unitree G1 人形（冻结模型，Dex3 三指手），自由行走Two Unitree G1 humanoids (frozen model, Dex3 three-finger hands), walking freely
是否拒答Refusal: 否No
见证结果Witness result: 成功：选定运行先用约 75 s 清障并返回，再在同一次运行中完成搬运与交付，仿真总用时 259.74 s；同一冻结源码预声明的三次运行全部通过，均为 259.74 s。仅证明本配置与记录环境中的复现，不作随机鲁棒性或第三台机器人收益结论。Success: the selected run spent about 75 s clearing the obstacle and returning, then carried and delivered the box in the same continuous run, 259.74 s of simulated time in total; all three predeclared runs of the same frozen source passed, each in 259.74 s. This shows repeatability for this configuration and recorded environment only, not randomized robustness or a benefit from a third robot.

8×

8 倍速 · 原始 258 s，播放 32 s8x speed · original 258 s, plays in 32 s

M2-6三台 G1 分工清障与搬运Three G1s split clearing and carrying

Unitree G1执行任务Solve the task

条件Conditions: 与 M2-5 相同的物体布局、0.5 kg 负载、障碍和暂存区，原两台 G1 初态不变，另加一台相同版本 G1 及合法起始站位。agent 可另选合法分工，也可让第三台等待，不要求三台同时抬箱；固定实例，专家使用仿真状态反馈。Same object layout, 0.5 kg load, obstacle and staging area as M2-5, with the original two G1s unchanged plus a third G1 of the same version at a legal start pose. An agent may choose another legal division of labour or let the third robot wait, and all three need not lift the box; fixed instance, and the expert uses simulator state feedback.
任务设计Task design: 专家让第三台 G1 负责把障碍推入暂存区、撤离并稳停，原两台执行双端起箱、转弯运输、下放与释放；实现仍记录等待与交接，不假定清障与运输可任意并行。在 M2-5 终态基础上增加第三台的稳定和接触检查，并比较成功率、总时间、动作成本、等待与相互阻挡。The expert assigns the third G1 to push the obstacle into the staging area, withdraw and stand still, while the original two do the two-ended lift, turning carry, lowering and release; waits and handoffs are still recorded rather than assuming clearing and carrying can run in parallel at will. Stability and contact checks for the third robot are added to the M2-5 terminal conditions, and the comparison covers success rate, total time, action cost, waiting and mutual blocking.
工具Tools: 可移动障碍（0.35 × 0.35 × 0.45 m，0.6 kg，不是固定墙）Movable obstacle (0.35 × 0.35 × 0.45 m, 0.6 kg; not a fixed wall)地面暂存区（0.50 × 0.55 m），清障后障碍须稳定留在其中Floor staging area (0.50 × 0.55 m) where the cleared obstacle must come to rest
Embodiment: 三台相同版本的 Unitree G1 人形（Dex3 三指手）Three Unitree G1 humanoids of the same version (Dex3 three-finger hands)
是否拒答Refusal: 否No
见证结果Witness result: 成功：选定运行一次连续完成清障、搬运、交付与三机停稳，仿真用时 257.96 s。同一冻结源码共 6 次运行，3 次成功、3 次失败（1 次未达到终态保持，2 次未完成有效运输）；后续指定节点组的 3 次完整运行均通过。不证明跨环境稳定，也不能由时间差推断第三台机器人带来效率收益。Success: the selected run cleared the obstacle, carried and delivered the box and brought all three robots to rest in one continuous run, 257.96 s of simulated time. The same frozen source was run 6 times with 3 passes and 3 failures (one missed the terminal hold, two did not complete effective transport); the three later runs on the selected node family all passed. This does not establish cross-environment stability, and no efficiency benefit of the third robot can be inferred from the time difference.