robo-si / SK

SK · Sokoban：双 G1 隐藏地图推箱清路Sokoban: two-G1 hidden-map box clearing

两台 G1 在未知的分室地图中探索，只凭各自第一视角 RGB、本体感知、自身目标标识和收到的显式消息构建并更新全局地图，通过物理推箱清理通路，最终同时稳定驻留在各自的目标区；箱子没有交付目标，只能推，不能抓、拉、抬。核心问题是能否从局部观察构建一致、可更新的全局地图并据此行动，以及显式通信（C+）相对无显式通信（C−）是否改善建图与执行。软件工具可由 agent 在私有沙盒内按需编写，实体工具不是通关要求。Two G1 humanoids explore an unknown divided-room map and, using only their own first-person RGB, proprioception, own goal identity and explicitly received messages, build and update a global map, physically push crates to clear routes, and finally both hold steady in their own goal zones; crates have no delivery targets and may only be pushed, never grasped, pulled or lifted. The core question is whether robots can build a consistent, updatable global map from local observations and act on it, and whether explicit communication (C+) improves mapping and execution over no explicit communication (C−). Agents may write software tools in a private sandbox as needed; physical tools are not required.

验收范围Accepted scope

作者已签收 6 个冻结的显式通信（C+）配置，依据第六轮冻结资格批次：8 个必需成功槽位（6 个配置见证与两次 L1 同种子复现）全部完整通过，6 份冻结见证包均已核对文件哈希并复算全部审计。6 个同实例无通信（C−）消融均在 650 s 时限内未完成，不算成功；L1 两个独立种子也完整通过，两张预留拓扑测试图一张因机器人碰撞失败、一张完整通过，仅作描述性证据。旧机器证据仍为 task_success=null、formal_admission=false，人工全视频审阅与独立人员复现未签名，前五轮资格批次的失败记录全部保留；作者构造的开发地图不能证明未见拓扑泛化或通信的总体收益。The author signed off the six frozen explicit-communication (C+) configurations on the basis of the sixth frozen qualification cohort, in which all 8 required-success slots (the six configuration witnesses and two same-seed L1 reproductions) passed in full and all six frozen witness packets had their file hashes checked and every audit recomputed. The six matched no-communication (C−) ablations all ran out the 650 s budget and do not count as successes; two independent L1 seeds also passed in full, while of two reserved-topology test maps one failed with a robot collision and one passed, as descriptive evidence only. The old machine evidence still records task_success=null and formal_admission=false, human full-video review and independent-person reproduction are unsigned, and the failures of the five earlier qualification cohorts are retained; the authored development maps do not demonstrate unseen-topology generalization or an overall communication benefit.

变化的条件轴Condition axes

地图拓扑与机器人能力两条轴：L1 为固定四箱开发图；L2 为三种作者构造的拓扑（信息互补、共享瓶颈、动态更新）；L3 沿用 L1 地图，仅将一台 G1 的手臂力矩上限降为 0.75 倍（A/B 互换）。签收的 6 个配置均为显式通信开启（C+），使用同一冻结策略、物理技能与 650 s 预算；C− 消融只作配对对照。Two axes, map topology and robot capability: L1 is the fixed four-crate development map; L2 has three authored topologies (complementary information, shared bottleneck, dynamic update); L3 keeps the L1 map and only scales one G1's arm effort limit to 0.75 (A/B swapped). All six accepted configurations have explicit communication enabled (C+) with the same frozen policy, physical skills and 650 s budget; the C− ablations serve only as paired controls.

Embodiments

两台同型 G1 人形机器人，每台每控制步提交 43 个绝对关节目标Two identical G1 humanoids, each submitting 43 absolute joint targets per control step
两台 G1，其中一台手臂力矩上限为 0.75 倍（L3 能力反转）Two G1s, one with its arm effort limit scaled to 0.75 (L3 capability reversal)

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：308 个文件，2.0 GiB。Reference evidence for 6/6 accepted configurations is archived: 308 files, 2.0 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.

SK-L1 SK-L2-INF SK-L2-BTL SK-L2-UPD SK-L3-A SK-L3-B

8×

8 倍速 · 原始 425 s，播放 53 s8x speed · original 425 s, plays in 53 s

SK-L1L1 · 固定四箱开发图L1 · Fixed four-crate development map

Unitree G1执行任务Solve the task

条件Conditions: SK-L1；两台同型 G1；C+ 显式通信（每条消息 ≤256 UTF-8 字节，每接收者最多 16 条待处理）；完整地图、箱子真值位姿和同伴隐藏状态均不提供；两机策略隔离执行（各自私有文件系统、进程与网络命名空间）；预算 650 s；成功＝两机同时站立并稳定驻留在各自目标区。SK-L1; two identical G1s; C+ explicit messages (at most 256 UTF-8 bytes each, at most 16 pending per recipient); no full map, ground-truth crate poses or hidden peer state; the two policies run isolated (private filesystem, process and network namespaces); 650 s budget; success = both robots standing and holding steady in their own goal zones at the same time.
任务设计Task design: 9×7 格（格距 0.9 m，含外墙）的双室地图，两机分别从两侧房间出发，须到达对面房间中带颜色标识的自身目标；四个可推动箱子阻挡通路，其中一个堵住共享门洞。机器人须从自身 RGB 重建地图、交换地图信息，并决定由谁推哪个箱子以及通行顺序。A two-room map of 9 × 7 cells (0.9 m grid, including outer walls); each robot starts in one room and must reach its own color-coded goal in the opposite room, while four pushable crates obstruct the routes, one of them blocking the shared doorway. The robots must rebuild the map from their own RGB, exchange map information, and decide who pushes which crate and in what order they pass.
工具Tools: 4 个可推动箱子（0.65×0.65×0.55 m，4 kg，摩擦系数 0.4），只能推，不能抓、拉、抬Four pushable crates (0.65 × 0.65 × 0.55 m, 4 kg, friction 0.4); push only, no grasping, pulling or lifting同伴 G1 与受限显式消息通道The peer G1 and a bounded explicit message channel可选：在私有沙盒内自行编写建图、地图融合与规划程序Optional: self-written mapping, map-fusion and planning programs in a private sandbox
Embodiment: 两台同型 G1 人形机器人Two identical G1 humanoids
是否拒答Refusal: 否No
见证结果Witness result: 冻结 C+ 见证成功：424.66 s 仿真时间内两机同时到达并驻留在各自目标区，30 个高层动作（3 次推箱）对应 79 个低层段，全部自动物理与证据检查通过；两机局部地图平均覆盖 89.7%，连通性问答全部正确。另有两次同种子复现和两个独立种子均完整通过；同实例 C− 消融在 650 s 时限内未完成。人工全视频审阅未签名。Frozen C+ witness succeeded: both robots reached and held their own goal zones within 424.66 s of simulated time, with 30 high-level actions (3 pushes) mapped to 79 low-level segments and every automated physics and evidence check passed; the two local maps averaged 89.7% coverage and answered every connectivity query correctly. Two same-seed reproductions and two independent seeds also passed in full; the matched C− ablation did not finish within 650 s. Human full-video review is unsigned.

8×

8 倍速 · 原始 456 s，播放 57 s8x speed · original 456 s, plays in 57 s

SK-L2-INFL2 · 信息互补L2 · Complementary information

Unitree G1执行任务Solve the task

条件Conditions: SK-L2；两台同型 G1；C+ 显式通信（限制同 L1）；原生执行；地图与箱子真值隐藏；预算 650 s；成功＝两机同时站立并稳定驻留在各自目标区。SK-L2; two identical G1s; C+ explicit messages (same limits as L1); native execution; map and crate ground truth hidden; 650 s budget; success = both robots standing and holding steady in their own goal zones at the same time.
任务设计Task design: 在双室布局上加入内凹墙段，形成非对称的侧室遮挡，两机初始视野覆盖不同分支；一方发现的死路或通道可帮助另一方减少重复探索。通信是否有益是实验问题，不预设结论。Inset wall segments added to the two-room layout create asymmetric side-room occlusion, so the two robots initially see different branches; a dead end or passage found by one robot can spare the other repeated exploration. Whether communication helps is an experimental question, not a presupposition.
工具Tools: 4 个可推动箱子（0.65×0.65×0.55 m，4 kg，摩擦系数 0.4），只能推，不能抓、拉、抬Four pushable crates (0.65 × 0.65 × 0.55 m, 4 kg, friction 0.4); push only, no grasping, pulling or lifting同伴 G1 与受限显式消息通道The peer G1 and a bounded explicit message channel可选：在私有沙盒内自行编写建图、地图融合与规划程序Optional: self-written mapping, map-fusion and planning programs in a private sandbox
Embodiment: 两台同型 G1 人形机器人Two identical G1 humanoids
是否拒答Refusal: 否No
见证结果Witness result: 冻结 C+ 见证成功：456.38 s 内两机同时到达并驻留，32 个高层动作（3 次推箱）对应 82 个低层段，全部自动检查通过；两机局部地图平均覆盖 86.5%，已回答的连通性问答全部正确（812 问中回答 765）。同实例 C− 消融在 650 s 时限内未完成。Frozen C+ witness succeeded: both robots reached and held their goals within 456.38 s, with 32 high-level actions (3 pushes) mapped to 82 low-level segments and every automated check passed; the local maps averaged 86.5% coverage and every answered connectivity query was correct (765 of 812 answered). The matched C− ablation did not finish within 650 s.

8×

8 倍速 · 原始 323 s，播放 40 s8x speed · original 323 s, plays in 40 s

SK-L2-BTLL2 · 共享瓶颈L2 · Shared bottleneck

Unitree G1执行任务Solve the task

条件Conditions: SK-L2（修订 2）；两台同型 G1；C+ 显式通信（限制同 L1）；原生执行；地图与箱子真值隐藏；预算 650 s；成功＝两机同时站立并稳定驻留在各自目标区。SK-L2 (revision 2); two identical G1s; C+ explicit messages (same limits as L1); native execution; map and crate ground truth hidden; 650 s budget; success = both robots standing and holding steady in their own goal zones at the same time.
任务设计Task design: 修订 2 用箱子同时堵住两条跨室通道（修订 1 存在无需推箱的绕行路线，保留为对照）；一方清路会改变另一方的可达区域，需要更新地图并协调执行顺序。不据此声称任务必须依赖推力。Revision 2 blocks both cross-room passages with crates (revision 1 allowed a valid no-push detour and is kept as a control); clearing a route by one robot changes what the other can reach, so the map and execution order must be updated. No claim is made that the task depends on pushing force.
工具Tools: 4 个可推动箱子（0.65×0.65×0.55 m，4 kg，摩擦系数 0.4），只能推，不能抓、拉、抬Four pushable crates (0.65 × 0.65 × 0.55 m, 4 kg, friction 0.4); push only, no grasping, pulling or lifting同伴 G1 与受限显式消息通道The peer G1 and a bounded explicit message channel可选：在私有沙盒内自行编写建图、地图融合与规划程序Optional: self-written mapping, map-fusion and planning programs in a private sandbox
Embodiment: 两台同型 G1 人形机器人Two identical G1 humanoids
是否拒答Refusal: 否No
见证结果Witness result: 冻结 C+ 见证成功：322.74 s 内两机同时到达并驻留，22 个高层动作（2 次推箱）对应 60 个低层段，全部自动检查通过；两机局部地图平均覆盖 77.8%，已回答的连通性问答全部正确（992 问中回答 760）。同实例 C− 消融在 650 s 时限内未完成。Frozen C+ witness succeeded: both robots reached and held their goals within 322.74 s, with 22 high-level actions (2 pushes) mapped to 60 low-level segments and every automated check passed; the local maps averaged 77.8% coverage and every answered connectivity query was correct (760 of 992 answered). The matched C− ablation did not finish within 650 s.

8×

8 倍速 · 原始 385 s，播放 48 s8x speed · original 385 s, plays in 48 s

SK-L2-UPDL2 · 动态更新L2 · Dynamic map update

Unitree G1执行任务Solve the task

条件Conditions: SK-L2；两台同型 G1；C+ 显式通信（限制同 L1）；原生执行；地图与箱子真值隐藏；预算 650 s；成功＝两机同时站立并稳定驻留在各自目标区。SK-L2; two identical G1s; C+ explicit messages (same limits as L1); native execution; map and crate ground truth hidden; 650 s budget; success = both robots standing and holding steady in their own goal zones at the same time.
任务设计Task design: 改变箱子停放位置（一个箱子改停在隔墙右侧）并设置非对称通道，推箱后通路改变，机器人须重新观察、更新地图并重规划。只有实际记录到失败与恢复时才计为恢复。Crate parking is changed (one crate now sits beside the right side of the partition) and an asymmetric passage is added, so routes change after pushing and the robots must re-observe, update the map and replan. Recovery is credited only when a failure and its recovery are actually recorded.
工具Tools: 4 个可推动箱子（0.65×0.65×0.55 m，4 kg，摩擦系数 0.4），只能推，不能抓、拉、抬Four pushable crates (0.65 × 0.65 × 0.55 m, 4 kg, friction 0.4); push only, no grasping, pulling or lifting同伴 G1 与受限显式消息通道The peer G1 and a bounded explicit message channel可选：在私有沙盒内自行编写建图、地图融合与规划程序Optional: self-written mapping, map-fusion and planning programs in a private sandbox
Embodiment: 两台同型 G1 人形机器人Two identical G1 humanoids
是否拒答Refusal: 否No
见证结果Witness result: 冻结 C+ 见证成功：384.66 s 内两机同时到达并驻留，28 个高层动作（3 次推箱）对应 70 个低层段，全部自动检查通过；3 次箱格变化均被两机地图检出（中位延迟约 24.4 s），地图平均覆盖 88.9%。同实例 C− 消融在 650 s 时限内未完成。Frozen C+ witness succeeded: both robots reached and held their goals within 384.66 s, with 28 high-level actions (3 pushes) mapped to 70 low-level segments and every automated check passed; all 3 crate cell changes were detected in both robots' maps (median latency about 24.4 s) and the maps averaged 88.9% coverage. The matched C− ablation did not finish within 650 s.

8×

8 倍速 · 原始 407 s，播放 51 s8x speed · original 407 s, plays in 51 s

SK-L3-AL3 · 能力反转 A（g1_a 手臂力矩降低）L3 · Capability reversal A (reduced arm effort on g1_a)

Unitree G1执行任务Solve the task

条件Conditions: SK-L3；沿用 L1 地图与目标；g1_a 手臂力矩上限为 0.75 倍，g1_b 保持 1.0；C+ 显式通信（限制同 L1）；原生执行；预算 650 s；成功＝两机同时站立并稳定驻留在各自目标区。SK-L3; same map and goals as L1; g1_a arm effort limit scaled to 0.75, g1_b kept at 1.0; C+ explicit messages (same limits as L1); native execution; 650 s budget; success = both robots standing and holding steady in their own goal zones at the same time.
任务设计Task design: 地图、箱子与目标均与 L1 相同，只改变已声明的手臂力矩能力，考察地图如何支持不同身体约束下的可达性判断与分工；与 L3-B 构成能力反转对。双 G1 力矩反转本身不能证明能力识别或角色适应。Map, crates and goals are identical to L1 and only the declared arm-effort capability changes, probing how the map supports reachability judgments and task division under different body limits; it forms a capability-reversal pair with L3-B. A dual-G1 effort reversal does not by itself demonstrate capability identification or role adaptation.
工具Tools: 4 个可推动箱子（0.65×0.65×0.55 m，4 kg，摩擦系数 0.4），只能推，不能抓、拉、抬Four pushable crates (0.65 × 0.65 × 0.55 m, 4 kg, friction 0.4); push only, no grasping, pulling or lifting同伴 G1 与受限显式消息通道The peer G1 and a bounded explicit message channel可选：在私有沙盒内自行编写建图、地图融合与规划程序Optional: self-written mapping, map-fusion and planning programs in a private sandbox
Embodiment: 两台 G1；g1_a 手臂力矩上限 0.75 倍Two G1s; g1_a arm effort limit ×0.75
是否拒答Refusal: 否No
见证结果Witness result: 冻结 C+ 见证成功：406.62 s 内两机同时到达并驻留，30 个高层动作（3 次推箱）对应 79 个低层段，全部自动检查通过；地图平均覆盖 89.7%，连通性问答全部正确。同实例 C− 消融在 650 s 时限内未完成；力矩差异由数值证据核对，录像本身不能体现。Frozen C+ witness succeeded: both robots reached and held their goals within 406.62 s, with 30 high-level actions (3 pushes) mapped to 79 low-level segments and every automated check passed; the maps averaged 89.7% coverage and answered every connectivity query correctly. The matched C− ablation did not finish within 650 s; the effort asymmetry is checked from numeric evidence and is not visible in the video itself.

8×

8 倍速 · 原始 443 s，播放 55 s8x speed · original 443 s, plays in 55 s

SK-L3-BL3 · 能力反转 B（g1_b 手臂力矩降低）L3 · Capability reversal B (reduced arm effort on g1_b)

Unitree G1执行任务Solve the task

条件Conditions: SK-L3；沿用 L1 地图与目标；g1_b 手臂力矩上限为 0.75 倍，g1_a 保持 1.0；C+ 显式通信（限制同 L1）；原生执行；预算 650 s；成功＝两机同时站立并稳定驻留在各自目标区。SK-L3; same map and goals as L1; g1_b arm effort limit scaled to 0.75, g1_a kept at 1.0; C+ explicit messages (same limits as L1); native execution; 650 s budget; success = both robots standing and holding steady in their own goal zones at the same time.
任务设计Task design: 与 L3-A 相同的地图与目标，但把力矩限制换到 g1_b，检验能力差异反转后可达性判断与分工是否随之调整；双 G1 力矩反转本身不能证明能力识别或角色适应。Same map and goals as L3-A but with the effort limit moved to g1_b, testing whether reachability judgments and task division follow when the capability difference is reversed; a dual-G1 effort reversal does not by itself demonstrate capability identification or role adaptation.
工具Tools: 4 个可推动箱子（0.65×0.65×0.55 m，4 kg，摩擦系数 0.4），只能推，不能抓、拉、抬Four pushable crates (0.65 × 0.65 × 0.55 m, 4 kg, friction 0.4); push only, no grasping, pulling or lifting同伴 G1 与受限显式消息通道The peer G1 and a bounded explicit message channel可选：在私有沙盒内自行编写建图、地图融合与规划程序Optional: self-written mapping, map-fusion and planning programs in a private sandbox
Embodiment: 两台 G1；g1_b 手臂力矩上限 0.75 倍Two G1s; g1_b arm effort limit ×0.75
是否拒答Refusal: 否No
见证结果Witness result: 冻结 C+ 见证成功：442.80 s 内两机同时到达并驻留，30 个高层动作（3 次推箱）对应 79 个低层段，全部自动检查通过；地图平均覆盖 89.7%，连通性问答全部正确。同实例 C− 消融在 650 s 时限内未完成；力矩差异由数值证据核对，录像本身不能体现。Frozen C+ witness succeeded: both robots reached and held their goals within 442.80 s, with 30 high-level actions (3 pushes) mapped to 79 low-level segments and every automated check passed; the maps averaged 89.7% coverage and answered every connectivity query correctly. The matched C− ablation did not finish within 650 s; the effort asymmetry is checked from numeric evidence and is not visible in the video itself.