robo-si / Hanoi

Hanoi · 桌面汉诺塔：跨身体工具使用Tabletop Tower of Hanoi: cross-embodiment tool use

固定骨盆的 G1 在桌面三杆汉诺塔场景中借助工具搬运带孔圆盘：每次只搬一个顶盘、大盘不得压小盘、整盘须越过杆顶并完全释放。十个独立任务按末端（Dex3、平行夹爪、Inspire 五指手）、能力公开或隐藏、夹持／腕力矩边界以及盘数分别登记，考察跨身体工具相性、能力边界判断与 self-modeling。场景提供 T1 侧夹钳与 T2 底部托叉，当前十项参考均以 T2 完成。A fixed-pelvis G1 moves annular disks between three rods on a table using tools: only one top disk moves at a time, a larger disk may never rest on a smaller one, and each disk must clear the rod top and be fully released. Ten separate tasks vary the end effector (Dex3, parallel gripper, Inspire five-finger hand), disclosed versus hidden capability, grip or wrist-torque boundaries, and disk count, testing cross-embodiment tool compatibility, capability-boundary judgment and self-modeling. The scene offers T1 tongs and a T2 under-disk fork; all ten current references use T2.

验收范围Accepted scope

2026-11-01 作者签收当前十项参考交付（10/10），各按冻结目标：01–04、06 为三盘七步整塔，05、09、10 为取具后最小盘 A→C 的单次完整转移，07/08 提交完整 15/31 步符号计划但只物理执行前三次转移。参考均为读取真值的特权专家解，不代表 learned self-modeling 成绩；任务族专项设计要求与正式准入仍未通过，T1 侧夹钳的完整路线相性仍未确定。On 2026-11-01 the author signed off the current ten-task reference delivery (10/10), each against its own frozen goal: Hanoi-01–04 and 06 are seven-move three-disk towers, 05, 09 and 10 are a single complete smallest-disk A→C transfer after tool pickup, and 07/08 submit complete 15/31-move symbolic plans but physically execute only the first three transfers. All references are privileged expert solutions, not learned self-modeling scores; the family-level design requirements and formal admission remain open, and full-route compatibility of the T1 tongs is still undetermined. (2026-11-01)

变化的条件轴Condition axes

末端身体（Dex3／平行夹爪／Inspire 五指手）；能力信息公开或隐藏（01↔03、02↔04、09↔10 为同一物理实例配对）；能力边界（夹持 0.25 倍、腕力矩 0.0625 倍）；盘数与目标范围（三盘整塔、四／五盘完整规划加前三步物理执行、单次转移）；时限 2400／4200／8500 仿真秒。End-effector body (Dex3 / parallel gripper / Inspire five-finger hand); disclosed versus hidden capability (01↔03, 02↔04 and 09↔10 are paired on the same physical instance); capability boundaries (grip at 0.25×, wrist torque at 0.0625×); disk count and goal scope (three-disk full tower, four/five-disk full plan with the first three physical transfers, single transfer); budgets of 2400/4200/8500 simulated seconds.

Embodiments

G1＋Dex3 三指灵巧手（固定骨盆）G1 + Dex3 three-finger hands (fixed pelvis)
G1＋合成平行夹爪（每侧行程 35 mm、25 N 上限，仿真设计，非商业硬件；固定骨盆）G1 + synthetic parallel gripper (35 mm stroke and 25 N per jaw; a simulated design, not commercial hardware; fixed pelvis)
G1＋Inspire 五指灵巧手（Unitree G1 配套资产；固定骨盆）G1 + Inspire five-finger hands (Unitree G1 companion asset; fixed pelvis)

统计Counts

10 个已验收任务／配置，其中 0 个拒答；10 个有倍速视频。10 accepted tasks/configurations, 0 of them refusal cases; 10 with sped-up videos.

证据备份Evidence backup

本族 10/10 个已验收配置的参考证据已归档：13,148 个文件，26.1 GiB。Reference evidence for 10/10 accepted configurations is archived: 13,148 files, 26.1 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.

Hanoi-01 Hanoi-02 Hanoi-03 Hanoi-04 Hanoi-05 Hanoi-06 Hanoi-07 Hanoi-08 Hanoi-09 Hanoi-10

8×

8 倍速 · 原始 25.5 min，播放 191 s8x speed · original 25.5 min, plays in 191 s

Hanoi-01Dex3·三盘整塔·能力公开Dex3, three-disk full tower, capability disclosed

Unitree G1执行任务Solve the task

条件Conditions: 三盘、标称能力，公开本实例的有效能力参数；固定骨盆 G1＋Dex3；T1、T2 两种工具在场；目标为三盘七步整塔 A→C；时限 2400 仿真秒。Three disks, nominal capability, with this instance's effective capability parameters disclosed; fixed-pelvis G1 + Dex3; tools T1 and T2 present; goal is the seven-move three-disk tower A→C; 2400 simulated-second budget.
任务设计Task design: 桌面三根 0.18 m 高的杆相距 0.25 m，三只 0.12 kg 带孔圆盘叠在 A 杆。机器人须自选工具、握点与单双手方式，每次只搬一个顶盘、越过杆顶、合法套杆并完全释放，最终整塔在 C 杆稳定 2 秒；测已知身体条件下的工具相性。Three 0.18 m rods stand 0.25 m apart on the table, with three 0.12 kg annular disks stacked on rod A. The robot chooses the tool, grasp point and one- or two-hand mode, moves one top disk at a time over the rod top, places it legally and fully releases it, and the full tower must stand stable on rod C for 2 s; this tests tool compatibility under known body conditions.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋Dex3 三指灵巧手（固定骨盆）G1 + Dex3 three-finger hands (fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a177 从初始状态连续完成 7/7 合法移动，用时 1528.88 仿真秒；全部 305,776 个物理步的独立回放误差为零。路线使用 T2 托叉并含三次桌面支撑重握；这是读取真值的特权专家参考，不是 self-modeling 成绩。Success: reference run a177 completed 7/7 legal moves continuously from the initial state in 1528.88 simulated seconds; an independent replay of all 305,776 physics steps matched with zero error. The route uses the T2 fork with three table-supported regrasps; this is a privileged expert reference, not a self-modeling score.

8×

8 倍速 · 原始 22.1 min，播放 166 s8x speed · original 22.1 min, plays in 166 s

Hanoi-02平行夹爪·三盘整塔·能力公开Parallel gripper, three-disk full tower, capability disclosed

Unitree G1执行任务Solve the task

条件Conditions: 与 01 相同的盘、杆和工具，只把末端换成平行夹爪并公开相应能力信息；三盘七步整塔；时限 2400 仿真秒。Same disks, rods and tools as Hanoi-01; only the end effector is swapped for the parallel gripper, with its capability information disclosed; seven-move three-disk tower; 2400 simulated-second budget.
任务设计Task design: 测换身体后是否正确调整工具或操作路线；与 01 比较共同操作。Tests whether the tool or operating route is adapted correctly after the body changes; compared with Hanoi-01 on the shared operation.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋合成平行夹爪（每侧行程 35 mm、25 N 上限，仿真设计，非商业硬件；固定骨盆）G1 + synthetic parallel gripper (35 mm stroke and 25 N per jaw; a simulated design, not commercial hardware; fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a108 完成 7/7 合法移动，用时 1327.39 仿真秒；265,477 步完整独立回放误差为零。平行夹爪路线无需重握即完成与 01 相同的高层移动，这是观察到的差异，不声明因果必要性。Success: reference run a108 completed 7/7 legal moves in 1327.39 simulated seconds; the full 265,477-step independent replay matched with zero error. The gripper route completes the same high-level moves as Hanoi-01 without regrasping; this is an observed difference, not a claim of causal necessity.

8×

8 倍速 · 原始 25.5 min，播放 191 s8x speed · original 25.5 min, plays in 191 s

Hanoi-03Dex3·三盘整塔·能力隐藏Dex3, three-disk full tower, capability hidden

Unitree G1执行任务Solve the task

条件Conditions: 与 01 同一物理实例，只隐藏所选有效能力参数；三盘七步整塔；有限探测预算；时限 2400 仿真秒。Same physical instance as Hanoi-01, with only the selected effective capability parameters hidden; seven-move three-disk tower; bounded probe budget; 2400 simulated-second budget.
任务设计Task design: agent 须先做有限探测、预测工具与路线可行性，再完成整塔；与 01 配对，测自身能力估计。The agent must run bounded probes and predict tool and route feasibility before completing the tower; paired with Hanoi-01 to test self-capability estimation.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋Dex3 三指灵巧手（固定骨盆）G1 + Dex3 three-finger hands (fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a184 与 01 使用同一冻结源码独立执行，完成 7/7，用时 1528.88 仿真秒，305,776 步完整回放误差为零。特权专家不经过探测与预测，隐藏条件下的 self-modeling 尚未评测。Success: reference run a184 executed independently with the same frozen source as Hanoi-01, completing 7/7 in 1528.88 simulated seconds; the full 305,776-step replay matched with zero error. The privileged expert does not probe or predict, so self-modeling under the hidden condition is not yet evaluated.

8×

8 倍速 · 原始 22.1 min，播放 166 s8x speed · original 22.1 min, plays in 166 s

Hanoi-04平行夹爪·三盘整塔·能力隐藏Parallel gripper, three-disk full tower, capability hidden

Unitree G1执行任务Solve the task

条件Conditions: 与 02 同一物理实例，只隐藏能力信息；三盘七步整塔；有限探测预算；时限 2400 仿真秒。Same physical instance as Hanoi-02, with only the capability information hidden; seven-move three-disk tower; bounded probe budget; 2400 simulated-second budget.
任务设计Task design: 采用与 03 相同的探测与预测协议，测平行夹爪下的 self-modeling 与工具选择。Uses the same probe-and-predict protocol as Hanoi-03 to test self-modeling and tool choice with the parallel gripper.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋合成平行夹爪（每侧行程 35 mm、25 N 上限，仿真设计，非商业硬件；固定骨盆）G1 + synthetic parallel gripper (35 mm stroke and 25 N per jaw; a simulated design, not commercial hardware; fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：a108 配对独立运行完成 7/7，用时 1327.39 仿真秒，完整回放误差为零。特权专家不经过探测与预测，self-modeling 尚未评测。Success: the paired independent a108 run completed 7/7 in 1327.39 simulated seconds, and the full replay matched with zero error. The privileged expert does not probe or predict, so self-modeling is not yet evaluated.

8×

8 倍速 · 原始 211 s，播放 26 s8x speed · original 211 s, plays in 26 s

Hanoi-05Dex3·夹持边界·单次完整转移Dex3, grip boundary, single complete transfer

Unitree G1执行任务Solve the task

条件Conditions: 夹持能力降为标称的 0.25 倍且不公开；授权修订 single-transfer-v2：保留三盘场景，取具后只把最小盘 A→C、撤具，其余两盘留在 A，全部盘独立稳定 2 秒；时限 2400 仿真秒。Grip capability reduced to 0.25 of nominal and not disclosed; authorized revision single-transfer-v2 keeps the three-disk scene but requires only taking a tool, moving the smallest disk A→C and withdrawing the tool, with the other two disks left on A and every disk independently stable for 2 s; 2400 simulated-second budget.
任务设计Task design: 该能力下原工具／握点组合失效，而同能力下调整握法仍能完成，需通过探测发现边界并改换路线，且不公开“夹持下降”标签。原七步整塔版本多轮完成五次放盘后，失败集中在空叉反复换握，经作者允许收敛为一次完整转移。At this capability the original tool/grasp combination fails while an adjusted grasp at the same capability still works, so the boundary must be found by probing and the route changed; the reduced-grip label is never disclosed. The original seven-move version reached five placements but kept failing during repeated empty-fork regrasps, so with the author's permission it was narrowed to one complete transfer.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋Dex3 三指灵巧手（固定骨盆）G1 + Dex3 three-finger hands (fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a307 完成 1/1 转移，用时 210.92 仿真秒，42,183 步独立回放误差为零，采用短握法替代路线。对照显示同一弱能力下原握法失败、较强能力下原握法成功；结论只针对这组控制器，不含重复换握维护。Success: reference run a307 completed the 1/1 transfer in 210.92 simulated seconds, and the 42,183-step independent replay matched with zero error; it uses a short-grasp alternative route. Controls show the original grasp fails at the same weak capability and succeeds at a stronger one; the conclusion covers these controllers only and excludes repeated regrasp maintenance.

8×

8 倍速 · 原始 26.3 min，播放 197 s8x speed · original 26.3 min, plays in 197 s

Hanoi-06平行夹爪·腕力矩边界·三盘整塔Parallel gripper, wrist-torque boundary, three-disk full tower

Unitree G1执行任务Solve the task

条件Conditions: 腕部力矩能力降为标称的 0.0625 倍且不公开（2026-10-30 标定后冻结）；三盘七步整塔；时限 2400 仿真秒。Wrist torque capability reduced to 0.0625 of nominal and not disclosed (frozen after calibration on 2026-10-30); seven-move three-disk tower; 2400 simulated-second budget.
任务设计Task design: 原工具／握点组合在该腕力矩下失效，但轻工具、短力臂或双手分担中至少一条路线仍可执行；测力臂选择与能力边界判断。The original tool/grasp combination fails at this wrist torque, but at least one route (lighter tool, shorter lever arm or two-hand load sharing) still works; this tests lever-arm choice and capability-boundary judgment.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋合成平行夹爪（每侧行程 35 mm、25 N 上限，仿真设计，非商业硬件；固定骨盆）G1 + synthetic parallel gripper (35 mm stroke and 25 N per jaw; a simulated design, not commercial hardware; fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a259 完成 7/7，用时 1578.90 仿真秒，315,780 步完整回放误差为零。对照显示同一固定握法在相同腕力矩上限下饱和失败，短握法与只增大腕力矩的对照均成功；不证明所有策略在该能力下都会失败。Success: reference run a259 completed 7/7 in 1578.90 simulated seconds, and the full 315,780-step replay matched with zero error. Controls show the same fixed grasp saturates and fails under the same wrist cap, while a short grasp and a wrist-torque-only increase both succeed; this does not prove every policy fails at this capability.

8×

8 倍速 · 原始 10.9 min，播放 81 s8x speed · original 10.9 min, plays in 81 s

Hanoi-07Dex3·四盘·完整规划＋前三步物理执行Dex3, four disks, full plan plus first three physical transfers

Unitree G1执行任务Solve the task

条件Conditions: 四盘、能力公开；授权修订 full-plan-three-transfer-v2：须提交完整 15 步整塔计划，只物理执行前三次转移（上方两盘移到 C，其余留 A）；时限 4200 仿真秒。Four disks, capability disclosed; authorized revision full-plan-three-transfer-v2 requires the complete 15-move tower plan but physically executes only its first three transfers (top two disks to C, the rest left on A); 4200 simulated-second budget.
任务设计Task design: 测符号规划深度与初段物理对应：实际三步须与提交计划的前三步一致，包含取具、一次真实桌面重握、越杆与稳定释放，最后全部盘独立稳定 2 秒。Tests symbolic planning depth and its match to the opening physical moves: the three executed moves must equal the plan's first three, including tool pickup, one real table-supported regrasp, rod clearance and stable release, after which every disk must stay independently stable for 2 s.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋Dex3 三指灵巧手（固定骨盆）G1 + Dex3 three-finger hands (fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a286 提交完整 15 步计划，完成 3/3 物理转移并真实重握，用时 651.58 仿真秒，130,316 步完整回放通过。只含前三次物理转移，其余 12 步仅为符号验证，不宣称整塔物理执行。Success: reference run a286 submitted the complete 15-move plan and completed 3/3 physical transfers with a real regrasp in 651.58 simulated seconds; the full 130,316-step replay passed. Only the first three physical transfers are included; the remaining 12 moves are symbolic only, and full-tower physical execution is not claimed.

8×

8 倍速 · 原始 10.3 min，播放 77 s8x speed · original 10.3 min, plays in 77 s

Hanoi-08Dex3·五盘·完整规划＋前三步物理执行Dex3, five disks, full plan plus first three physical transfers

Unitree G1执行任务Solve the task

条件Conditions: 五盘、能力公开；授权修订 full-plan-three-transfer-v2：须提交完整 31 步整塔计划，只物理执行前三次转移（上方两盘移到 B，其余留 A）；时限 8500 仿真秒。Five disks, capability disclosed; authorized revision full-plan-three-transfer-v2 requires the complete 31-move tower plan but physically executes only its first three transfers (top two disks to B, the rest left on A); 8500 simulated-second budget.
任务设计Task design: 与 07 对照完整符号规划深度，不以重复搬运次数衡量 reasoning；物理部分同样含取具、真实重握、越杆与稳定释放。Contrasts full symbolic planning depth with Hanoi-07 rather than measuring reasoning by repeated carries; the physical part again includes tool pickup, a real regrasp, rod clearance and stable release.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋Dex3 三指灵巧手（固定骨盆）G1 + Dex3 three-finger hands (fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a288 提交完整 31 步计划，完成 3/3 物理转移并真实重握，用时 617.89 仿真秒，123,578 步完整回放通过。只含前三次物理转移，其余 28 步仅为符号验证，不宣称整塔物理执行。Success: reference run a288 submitted the complete 31-move plan and completed 3/3 physical transfers with a real regrasp in 617.89 simulated seconds; the full 123,578-step replay passed. Only the first three physical transfers are included; the remaining 28 moves are symbolic only, and full-tower physical execution is not claimed.

8×

8 倍速 · 原始 281 s，播放 35 s8x speed · original 281 s, plays in 35 s

Hanoi-09五指手·单次转移·能力公开Five-finger hand, single transfer, capability disclosed

Unitree G1执行任务Solve the task

条件Conditions: 三盘初始场景、标称能力、公开；授权修订 single-transfer-v2：取具后把最小盘从 A 越杆搬到 C、撤具，其余两盘留在 A，全部盘稳定 2 秒；时限 2400 仿真秒。Three-disk initial scene, nominal capability, disclosed; authorized revision single-transfer-v2: take a tool, carry the smallest disk from A over the rod to C, withdraw the tool, leave the other two disks on A, and keep all disks stable for 2 s; 2400 simulated-second budget.
任务设计Task design: 测五指手的工具获取、承载和操作相性。旧七步候选多次完成首步，但重握后握内漂移、重复插入失败，因此收敛为一次完整工具转移；与 01/02 只比较共同首步。Tests tool acquisition, load bearing and operating compatibility of the five-finger hand. Earlier seven-move candidates repeatedly completed the first move but failed through in-hand drift after regrasps and repeated insertion, so the goal was narrowed to one complete tool transfer; comparison with Hanoi-01/02 uses only the shared first move.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋Inspire 五指灵巧手（Unitree G1 配套资产；固定骨盆）G1 + Inspire five-finger hands (Unitree G1 companion asset; fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a273 完成 1/1 转移，用时 280.74 仿真秒，56,147 步完整回放误差为零。Success: reference run a273 completed the 1/1 transfer in 280.74 simulated seconds, and the full 56,147-step replay matched with zero error.

8×

8 倍速 · 原始 281 s，播放 35 s8x speed · original 281 s, plays in 35 s

Hanoi-10五指手·单次转移·能力隐藏Five-finger hand, single transfer, capability hidden

Unitree G1执行任务Solve the task

条件Conditions: 与 09 同一三盘初始场景与物理条件，只隐藏能力信息；同为 single-transfer-v2 单次转移；有限探测预算；时限 2400 仿真秒。Same three-disk initial scene and physics as Hanoi-09, with only the capability information hidden; same single-transfer-v2 goal; bounded probe budget; 2400 simulated-second budget.
任务设计Task design: 有限探测后先预测所选工具／路线能否完成再执行；与 09 构成第三种身体的公开／隐藏配对。After bounded probing the agent predicts whether the chosen tool and route will work before executing; forms the disclosed/hidden pair with Hanoi-09 for the third body.
工具Tools: T1 侧夹钳：被动侧向夹持工具T1 tongs: passive side-clamping toolT2 底部托叉：从盘下支撑的托具（当前参考解使用）T2 fork: under-disk support tool (used by the current reference)
Embodiment: G1＋Inspire 五指灵巧手（Unitree G1 配套资产；固定骨盆）G1 + Inspire five-finger hands (Unitree G1 companion asset; fixed pelvis)
是否拒答Refusal: 否No
见证结果Witness result: 成功：参考运行 a274 与 09 同物理条件、同冻结源码独立完成 1/1，用时 280.74 仿真秒，56,147 步完整回放通过。特权专家不含探测与预测评测。Success: reference run a274, under the same physics and frozen source as Hanoi-09, independently completed the 1/1 transfer in 280.74 simulated seconds, and the full 56,147-step replay passed. The privileged expert does not include probe-and-predict evaluation.