robo-si / TT

TT · 乒乓对打与死球判断Table-tennis rallies and dead-ball judgment

以比赛场景为主体：同一颗球在两台机器人之间连续对打，不重置、不补发，对手只能经有界执行器物理击球，对面可以是固定参考对手或另一种身体（UR10e、G1、A3）。另设死球判断：合法球须回击，死球不得触碰并保持准备姿态。人形机器人不焊接、靠真实脚底接触站立，手持拍版本连球拍也不焊接；两个已批准的单次回球参考（UR10e、G1）作为原始任务冻结。The family is built around match play: one ball is rallied continuously between two robots with no resets or replacement balls, and the opponent may strike only through bounded actuators; the opponent is either a fixed reference paddle or another body (UR10e, G1, A3). Dead-ball tasks require returning legal balls while leaving dead balls untouched and holding the ready pose. Humanoids are unwelded and balance on real foot contacts, the hand-held variant does not weld the racket either, and the two approved single-return references (UR10e, G1) are frozen as the original tasks.

验收范围Accepted scope

签收记录：参考已批准，D2 明确验收；任务表以 D2 计 15/15 完成，扩展配置 46/47 有见证，TT-D-A 的 D1 档仍失败（有证据）。两个原始单次回球参考于 2026-10-27 获批正式发布（table-tennis-0.1.0）。对打均为合作模式，专家使用特权状态离线规划，未做留出 seed 泛化评测；他人复现与不可信智能体沙盒接入仍待完成。Sign-off: references approved and D2 explicitly accepted; counted at D2 the task table is 15/15 complete, 46/47 extension configurations have witnesses, and TT-D-A at tier D1 still fails (with evidence). The two original single-return references were approved for formal release on 2026-10-27 (table-tennis-0.1.0). All rallies are cooperative, experts plan offline from privileged state, and no held-out seed evaluation was run; independent-person reproduction and isolated untrusted-agent sandbox integration remain open.

变化的条件轴Condition axes

被测身体（UR10e、G1、A3、G1＋五指手）、对手（发球机、固定参考对手、同构或异构专家）与场景类型（单次回球、连续对打、死球判断、手持拍对打）。D1/D2/D3 是任务内难度档（对打 N=3/5/10），不另计任务；本页逐任务以 D2 结果计。Subject body (UR10e, G1, A3, G1 with a five-finger hand), opponent (launcher, fixed reference opponent, same-body or cross-body expert) and scene type (single return, continuous rally, dead-ball judgment, hand-held racket rally). D1/D2/D3 are difficulty tiers inside each task (rally N=3/5/10) and do not add tasks; each task here is counted by its D2 result.

Embodiments

UR10e 六关节固定基座机械臂，球拍刚性固连在法兰UR10e 6-joint fixed-base arm, racket rigidly mounted on the flange
Unitree G1，29 关节浮动基座，靠脚底接触站立，球拍被动固连在右腕Unitree G1, 29 joints, floating base on real foot contacts, racket passively mounted on the right wrist
A3 Ultra 人形（a3_ultra_t3d0），不焊接站立，31 个有界关节力矩，球拍固连腕部（偏置 0.22 m）A3 Ultra humanoid (a3_ultra_t3d0), unwelded, 31 bounded joint torques, racket mounted on the wrist (0.22 m offset)
G1＋Inspire 五指手，球拍为独立刚体，只靠手指接触握持G1 with an Inspire five-finger hand; the racket is a free rigid body held only by finger contact

统计Counts

15 个已验收任务／配置，其中 0 个拒答；3 个有倍速视频。15 accepted tasks/configurations, 0 of them refusal cases; 3 with sped-up videos.

证据备份Evidence backup

本族 15/15 个已验收配置的参考证据已归档：5,524 个文件，4.2 GiB。Reference evidence for 15/15 accepted configurations is archived: 5,524 files, 4.2 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.

TT-ORIG-U TT-ORIG-G TT-R-U-ref TT-R-G-ref TT-R-A-ref TT-R-U-U TT-R-G-G TT-R-U-G TT-R-G-U TT-R-A-G TT-R-G-A TT-D-U TT-D-G TT-D-A TT-R-GH-ref

1×

原速 · 原始 2.4 s，播放 2.4 sreal time · original 2.4 s, plays in 2.4 s

TT-ORIG-UUR10e 单次目标回球UR10e single target return

UR10e执行任务Solve the task

条件Conditions: 原始任务，固定基座 UR10e；发球机单发无旋球，四个公开固定配置（nominal、left、right、fast），每局 2.4 s；物理 2 kHz、参考 500 Hz，特权仿真状态观测。Original task with a fixed-base UR10e; one spin-free launcher feed in four public fixed configurations (nominal, left, right, fast), 2.4 s per episode; physics at 2 kHz, references at 500 Hz, privileged simulator state.
任务设计Task design: 来球在本方台面弹地一次后用球拍合法击回，过网落入对方半台的 0.30 m 目标圆，再回到准备姿态并保持。分层专家为准备接球 → 击球与随挥 → 回位保持，GPU 评估 28,033 个出球候选；评分为 stage-scoring/1：有效击球 25＋越网 25＋对侧落台 20＋命中目标 15＋回位保持 15。After one bounce on its own half the ball must be struck legally over the net into the 0.30 m target disk on the far half, then the arm returns to and holds its ready pose. The hierarchical expert runs prepare → strike and follow-through → recover and hold, with a GPU search over 28,033 outgoing-shot candidates; stage-scoring/1 awards contact 25 + net 25 + far-half landing 20 + target 15 + recovery 15.
工具Tools: 法兰固连球拍Flange-mounted racket发球机Ball launcher
Embodiment: UR10e 固定基座机械臂UR10e fixed-base arm
是否拒答Refusal: 否No
见证结果Witness result: 已批准参考：四个配置各 3 次共 12/12 完整成功，每条 100 分，新场景力矩回放 4/4；静止机械臂对照 25 分未成功，越界宽来球 0 分未执行。见证为 nominal 配置的一次完整运行；12 次是四个固定配置的重复，不代表随机泛化成功率。Approved reference: 12/12 full successes (three runs in each of four configurations), each scoring 100, with 4/4 fresh-scene torque replays; a stationary-arm control scores 25 without success and an out-of-bounds wide feed scores 0 and is not executed. The witness is one complete nominal run; the 12 runs are repeats of four fixed configurations, not a random-generalisation success rate.

1×

原速 · 原始 3.0 s，播放 3.0 sreal time · original 3.0 s, plays in 3.0 s

TT-ORIG-GG1 自由站立回球与稳定恢复G1 free-standing return and stable recovery

Unitree G1执行任务Solve the task

条件Conditions: 原始任务，G1 双脚原地站立，靠真实地面接触力，骨盆既不固定也不驱动；29 个关节力矩，500 Hz 浮动基座接触约束 QP，MuJoCo 2 kHz；四个公开固定配置（nominal、left、right、fast），每局 3.0 s；特权状态、无旋。Original task: G1 stands in place on real ground contact with an unfixed, unactuated pelvis; 29 joint torques via a 500 Hz floating-base contact-constrained QP, MuJoCo at 2 kHz; four public fixed configurations (nominal, left, right, fast), 3.0 s per episode; privileged state, no spin.
任务设计Task design: 与 UR10e 相同的单次目标回球，但挥拍与全身平衡耦合：恢复须双脚受力、躯干直立、根部与关节低速并持续保持，击球落点正确也不能掩盖失衡。120 g 球拍被动固连右腕；不含迈步或连续对打。The same single target return as the UR10e task, but the swing is coupled to whole-body balance: recovery requires both feet loaded, an upright torso and low root and joint speeds held over time, so a correct landing cannot hide a loss of balance. The 120 g racket is a passive wrist attachment; no stepping or continuous rally.
工具Tools: 右腕被动固连球拍（120 g）Racket passively mounted on the right wrist (120 g)发球机Ball launcher
Embodiment: Unitree G1 自由站立Unitree G1, free-standing
是否拒答Refusal: 否No
见证结果Witness result: 已批准参考：12/12 完整成功，每条 100 分，独立力矩回放 4/4；原地保持对照 25 分未成功，零力矩与不可达来球均为 0 分。nominal 配置中最小脚底力约 117 N、躯干倾斜约 0.2°、脚滑移不到 1 mm；12 次是四个固定配置的重复。Approved reference: 12/12 full successes, each scoring 100, with 4/4 independent torque replays; the hold-pose control scores 25 without success and the zero-torque and unreachable-feed controls score 0. In the nominal configuration the minimum foot force is about 117 N, torso tilt about 0.2 degrees and foot slip under 1 mm; the 12 runs are repeats of four fixed configurations.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-U-refUR10e 对参考对手连续对打UR10e rally against the reference opponent

UR10e执行任务Solve the task

条件Conditions: D2：N=5，参考对手左右交替回球；同一颗球不重置，首球为原始 nominal 发球，其后每球都由对手物理击出；每拍 3 s、整局 20 s（3N+5）；特权状态、无旋、合作模式。D2: N=5 with the reference opponent alternating sides; one ball, no resets: the first ball is the original nominal feed and every later ball is struck physically by the opponent; 3 s per exchange and 20 s per episode (3N+5); privileged state, no spin, cooperative mode.
任务设计Task design: 参考对手为抽象龙门球拍（三向平移＋偏航＋俯仰，电机、速度与加速度有界），所有被测方共用，作为控制变量。被测方须在非复位状态下逐拍重新规划、回击到对方目标圆并在拍间恢复，评分 80·C/N＋10·H/N＋10·R_end。参考对手回球区间以原始来球为中心，测连续性而非覆盖范围。The reference opponent is an abstract gantry paddle (three slides plus yaw and pitch, with bounded motors, speed and acceleration) shared by all subjects as a controlled variable. The subject must replan each shot from a non-reset state, return to the far target disk and recover between shots; score = 80·C/N + 10·H/N + 10·R_end. The opponent's return envelope is centred on the original feed, so the task tests continuity rather than reach.
工具Tools: 法兰固连球拍Flange-mounted racket参考对手（龙门球拍）Reference opponent (gantry paddle)
Embodiment: UR10e 固定基座机械臂UR10e fixed-base arm
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：一条连续运行得 100 分（C=5/5、H=5、R_end=1）。D1、D3 也为 100 分；D3 打满 10 拍后参考对手打出出界死球，被正确放过。D2 witness: one continuous run scores 100 (C=5/5, H=5, R_end=1). D1 and D3 also score 100; in D3, after ten returns the reference opponent plays a wide dead ball that is correctly let go.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-G-refG1 对参考对手连续对打G1 rally against the reference opponent

Unitree G1执行任务Solve the task

条件Conditions: D2：N=5，参考对手左右交替回球；G1 双脚静止支撑、不焊接，右手击球线对准球台中线；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋、合作模式。D2: N=5 with the reference opponent alternating sides; G1 in unwelded static double support, its right-hand hitting line on the table centre line; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin, cooperative mode.
任务设计Task design: 同 TT-R-U-ref，被测方换成浮动基座 G1，复用原始全身控制器。额外考察连续击球的反作用力在拍间累积、平衡余量是否下降；挥拍可配合腰部转动。As TT-R-U-ref, but the subject is the floating-base G1 reusing the original whole-body controller. It additionally tests whether strike reactions accumulate between shots and erode the balance margin; the swing may use the waist.
工具Tools: 右腕被动固连球拍Racket passively mounted on the right wrist参考对手（龙门球拍）Reference opponent (gantry paddle)
Embodiment: Unitree G1 自由站立Unitree G1, free-standing
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：100 分（C=5/5、H=5、R_end=1），使用加入瞄准修正的冻结 G1 专家（tag h：在私有仿真副本中预测击球误差后修正瞄准点，真实执行仍只经有界执行器）。D1、D3 也为 100 分；旧版 tag g 的 D1–D3 见证保留。D2 witness: 100 (C=5/5, H=5, R_end=1) with the frozen G1 expert tag h, which predicts its strike error in a private simulation copy and corrects the aim point while real execution still goes only through bounded actuators. D1 and D3 also score 100; the older tag-g D1–D3 witnesses are retained.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-A-refA3 对参考对手连续对打A3 rally against the reference opponent

A3执行任务Solve the task

条件Conditions: D2：N=5，参考对手左右交替回球；A3 Ultra 不焊接站立，31 个有界力矩、全身 QP 控制，球拍固连腕部（偏置 0.22 m），手指固定成拳；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋。D2: N=5 with the reference opponent alternating sides; unwelded A3 Ultra with 31 bounded torques under a whole-body QP, racket on the wrist (0.22 m offset), fingers fixed as a fist; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin.
任务设计Task design: 第二种人形对同一参考对手的基线，可与 TT-R-U-ref、TT-R-G-ref 直接比较。A3 身高、臂长与质量分布不同于 G1，平衡与挥拍的耦合方式也不同；击球可加入腰部偏航。A second-humanoid baseline against the same reference opponent, directly comparable with TT-R-U-ref and TT-R-G-ref. A3 differs from G1 in height, arm length and mass distribution, so balance couples differently with the swing; strikes may add waist yaw.
工具Tools: 腕部固连球拍Wrist-mounted racket参考对手（龙门球拍）Reference opponent (gantry paddle)
Embodiment: A3 Ultra 人形A3 Ultra humanoid
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：一条连续运行得 100 分（C=5/5、H=5、R_end=1），在线与离线裁判一致。D1、D3 也为 100 分；平衡余量随档位下降（D1 最小脚底力 211 N，D3 左脚低至 26 N），无跌倒。D2 witness: one continuous run scores 100 (C=5/5, H=5, R_end=1), with online and offline referees in agreement. D1 and D3 also score 100; the balance margin shrinks with tier (minimum foot force 211 N at D1, left foot down to 26 N at D3), with no fall.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-U-UUR10e 对 UR10e 对打UR10e versus UR10e rally

UR10e执行任务Solve the task

条件Conditions: D2：N=5，对手为冻结的 UR10e 连续对打专家（同一代码绕竖轴镜像），左右交替回球；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋、合作模式。D2: N=5 against a frozen UR10e rally expert (the same code mirrored about the vertical axis) that alternates sides; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin, cooperative mode.
任务设计Task design: 两台 UR10e 隔台对打，来球不再是预登记分布，而是同型机器人在同样身体限制下打出的球。除连续对打评分外，另报回合总长与失误归属。Two UR10e arms rally across the table, so the incoming balls are no longer a pre-registered distribution but shots produced by a same-type robot under the same body limits. Besides the rally score, total rally length and fault attribution are reported.
工具Tools: 法兰固连球拍Flange-mounted racketUR10e 专家对手UR10e expert opponent
Embodiment: UR10e 固定基座机械臂UR10e fixed-base arm
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：100 分（C=5/5、H=5、R_end=1）。D1、D3 也为 100 分；最后一拍后撤回的对手机械臂仍被球碰到，记为回合后接触，不计分。D2 witness: 100 (C=5/5, H=5, R_end=1). D1 and D3 also score 100; after the final return the withdrawn opponent arm is still hit by the ball, which is logged as a post-rally contact and not scored.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-G-GG1 对 G1 对打G1 versus G1 rally

Unitree G1执行任务Solve the task

条件Conditions: D2：N=5，两侧均为双脚静止支撑的 G1，对手为冻结的 G1 专家，左右交替回球；不允许迈步；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋、合作模式。D2: N=5 with a statically standing G1 on both sides, the opponent a frozen G1 expert alternating sides; no stepping; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin, cooperative mode.
任务设计Task design: 同构人形对打：双方都有平衡约束，回球误差相互放大，是本族风险最高的任务。来球偏离接球线时只能靠手臂加腰部够到。Same-body humanoid rally: both sides are balance-constrained and return errors amplify each other, making this the riskiest task in the family. When a ball drifts off the receiving line it can only be reached with the arm plus waist.
工具Tools: 右腕被动固连球拍Racket passively mounted on the right wristG1 专家对手G1 expert opponent
Embodiment: Unitree G1 自由站立Unitree G1, free-standing
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：100 分（C=5/5、H=5、R_end=1），双方均为加入瞄准修正的冻结专家 tag h，双脚保持原位。D1、D3 也为 100 分；旧版 tag g 的 D2 为 18 分、D3 为 9 分（G1 落点误差使下一球偏离接球线），失败记录保留。D2 witness: 100 (C=5/5, H=5, R_end=1), with both sides running the frozen aim-corrected expert tag h and feet kept in place. D1 and D3 also score 100; the older tag g scored 18 at D2 and 9 at D3 (G1 landing errors pushed the next ball off the receiver's line), and those failures are retained.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-U-GUR10e 对 G1 专家对打UR10e versus G1 expert rally

UR10e执行任务Solve the task

条件Conditions: D2：N=5，UR10e 为被测方，对手为冻结的 G1 专家（tag h），左右交替回球；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋、合作模式。D2: N=5 with UR10e as the subject and a frozen G1 expert (tag h) as the opponent alternating sides; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin, cooperative mode.
任务设计Task design: 异构对打：同一张球台，被测机械臂要接住另一种身体打出的回球，其分布由对面人形的平衡与够及限制决定。报告相对 TT-R-U-U 的 C 降幅作为跨身体适应量，不比较两种身体的力矩或能耗。Cross-body rally: on the same table the subject arm must handle returns from another body, whose distribution is shaped by the humanoid's balance and reach limits. The drop in C relative to TT-R-U-U is reported as cross-body adaptation; torques and energy of the two bodies are not compared.
工具Tools: 法兰固连球拍Flange-mounted racketG1 专家对手G1 expert opponent
Embodiment: UR10e 固定基座机械臂UR10e fixed-base arm
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：100 分（C=5/5、H=5、R_end=1）。D1、D3 也为 100 分；旧版 tag g 对手下 D1–D3 也有见证。D2 witness: 100 (C=5/5, H=5, R_end=1). D1 and D3 also score 100; D1–D3 were also witnessed against the older tag-g opponent.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-G-UG1 对 UR10e 专家对打G1 versus UR10e expert rally

Unitree G1执行任务Solve the task

条件Conditions: D2：N=5，G1 为被测方，对手为冻结的 UR10e 专家，左右交替回球；G1 双脚静止支撑；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋、合作模式。D2: N=5 with G1 as the subject in static double support and a frozen UR10e expert as the opponent alternating sides; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin, cooperative mode.
任务设计Task design: 与 TT-R-U-G 对调被测方：人形要接住机械臂打出的回球，报告相对 TT-R-G-G 的 C 降幅。两种身体的回球分布差异是待测假设，不预设结论。TT-R-U-G with the roles swapped: the humanoid must handle returns from the arm, and the drop in C relative to TT-R-G-G is reported. Differences between the two bodies' return distributions are a hypothesis under test, not an assumed conclusion.
工具Tools: 右腕被动固连球拍Racket passively mounted on the right wristUR10e 专家对手UR10e expert opponent
Embodiment: Unitree G1 自由站立Unitree G1, free-standing
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：100 分（C=5/5、H=5、R_end=1），被测 G1 为冻结专家 tag h。D1、D3 也为 100 分；旧版 tag g 的 D1–D3 见证保留，加入腰部候选前的开发尝试因身体触球失败，记录保留。D2 witness: 100 (C=5/5, H=5, R_end=1) with the subject G1 running the frozen expert tag h. D1 and D3 also score 100; the older tag-g D1–D3 witnesses are retained, as are earlier development attempts that failed on body contact before waist candidates were added.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-A-GA3 对 G1 专家对打A3 versus G1 expert rally

A3执行任务Solve the task

条件Conditions: D2：N=5，A3 为被测方，对手为冻结的 G1 专家（tag h），左右交替回球；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋、合作模式。D2: N=5 with A3 as the subject and a frozen G1 expert (tag h) as the opponent alternating sides; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin, cooperative mode.
任务设计Task design: 人形对人形且非镜像：两种人形的身高、臂长和平衡约束不同。报告相对 TT-R-A-ref 的 C 降幅。Humanoid against humanoid, not mirrored: the two humanoids differ in height, arm length and balance constraints. The drop in C relative to TT-R-A-ref is reported.
工具Tools: 腕部固连球拍Wrist-mounted racketG1 专家对手G1 expert opponent
Embodiment: A3 Ultra 人形A3 Ultra humanoid
是否拒答Refusal: 否No
见证结果Witness result: D2 见证（2026-10-30 重跑）：100 分（C=5/5、H=5、R_end=1），在线与离线裁判一致，相对 TT-R-A-ref 的 C 降幅为 0。D1、D3 也为 100 分；旧版 v2 的 D2 同为 100 分，其 D3 因 G1 对手失误只有 45 分，记录保留。D2 witness (re-run 2026-10-30): 100 (C=5/5, H=5, R_end=1), online and offline referees agree, C drop versus TT-R-A-ref is 0. D1 and D3 also score 100; the older v2 run also scored 100 at D2, while its D3 scored only 45 because of a G1 opponent fault, and that record is retained.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-G-AG1 对 A3 专家对打G1 versus A3 expert rally

Unitree G1执行任务Solve the task

条件Conditions: D2：N=5，G1 为被测方（冻结专家 tag h），对手为 A3 专家，左右交替回球；同一颗球不重置，每拍 3 s、整局 20 s；特权状态、无旋、合作模式。D2: N=5 with G1 as the subject (frozen expert tag h) and an A3 expert as the opponent alternating sides; one ball with no resets, 3 s per exchange and 20 s per episode; privileged state, no spin, cooperative mode.
任务设计Task design: 与 TT-R-A-G 对调被测方：G1 要接住 A3 打出的回球，报告相对 TT-R-G-G 的 C 降幅。TT-R-A-G with the roles swapped: G1 must handle returns struck by A3, and the drop in C relative to TT-R-G-G is reported.
工具Tools: 右腕被动固连球拍Racket passively mounted on the right wristA3 专家对手A3 expert opponent
Embodiment: Unitree G1 自由站立Unitree G1, free-standing
是否拒答Refusal: 否No
见证结果Witness result: D2 见证（2026-10-30 重跑）：100 分（C=5/5、H=5、R_end=1），相对 TT-R-G-G 的 C 降幅为 0。D1、D3 也为 100 分；A3 对手规划日志的坐标系标注以派生文件修正，原始文件、执行命令与裁判结果未改。D2 witness (re-run 2026-10-30): 100 (C=5/5, H=5, R_end=1), C drop versus TT-R-G-G is 0. D1 and D3 also score 100; a coordinate-frame label in the A3 opponent's plan logs is corrected in derived files, while original files, executed commands and referee results are unchanged.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-D-UUR10e 死球判断UR10e dead-ball judgment

UR10e执行任务Solve the task

条件Conditions: D2：一条连续运行中发球机依次发出 6 颗独立球，3 个为原始合法来球，3 个为擦网死球（擦网后落回对方一侧或落到台外）；每球周期 2.4 s（同原任务）；特权状态、无旋。D2: in one continuous run the launcher fires 6 separate balls, 3 original legal feeds and 3 net-clip dead balls (clipping the net and falling back on the far side or off the table); 2.4 s per ball as in the original task; privileged state, no spin.
任务设计Task design: 被测方逐球判断：合法球回击（按原 5 段评分），死球不得用球拍或身体触碰并保持准备姿态。死球按本任务约定：未触本方台面即越过端线或侧线、擦网后落回对方一侧或台外、擦台侧；死球误击与合法球误放分开统计。The subject judges every ball: legal balls are returned (scored by the original five stages) and dead balls must not be touched by racket or body while the ready pose is held. Dead balls follow the task convention (crossing the end or side line without touching the own half, net clips falling back or off the table, table-side hits); dead-ball strikes and false let-gos are counted separately.
工具Tools: 法兰固连球拍Flange-mounted racket发球机（6 球弹匣）Ball launcher (6-ball magazine)
Embodiment: UR10e 固定基座机械臂UR10e fixed-base arm
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：6 球全部判对，3 个合法球回击、3 个死球正确放过，死球误击 0、合法球误放 0，物理检查通过。D1、D3 也全部判对（三档共 18 球）。D2 witness: all 6 balls judged correctly, 3 legal balls returned and 3 dead balls correctly let go, with zero dead-ball strikes and zero false let-gos; physics checks pass. D1 and D3 are also all correct (18 balls over three tiers).

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-D-GG1 死球判断与躯干避让G1 dead-ball judgment with body avoidance

Unitree G1执行任务Solve the task

条件Conditions: D2：一条连续运行发出 6 颗独立球，3 个原始合法来球、3 个擦网死球（落回对方一侧或台外）；每球周期 3.0 s（同原任务）；G1 双脚静止支撑；特权状态、无旋。D2: one continuous run with 6 separate balls, 3 original legal feeds and 3 net-clip dead balls (falling back on the far side or off the table); 3.0 s per ball as in the original task; G1 in static double support; privileged state, no spin.
任务设计Task design: 同 TT-D-U，但 G1 躯干本身在球路上，必要时须主动避让（腰部侧倾、俯仰等），身体不得触碰死球，同时不得摔倒。As TT-D-U, but the G1 torso itself sits in the ball path and must actively avoid when needed (waist lean and pitch), keeping the body off dead balls without falling.
工具Tools: 右腕被动固连球拍Racket passively mounted on the right wrist发球机（6 球弹匣）Ball launcher (6-ball magazine)
Embodiment: Unitree G1 自由站立Unitree G1, free-standing
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：6 球全部判对，3 个合法球回击、3 个死球正确放过，物理检查通过。D1、D3 也全部判对（三档共 18 球）；D1 中瞄向头部与躯干的死球靠弯腰避让，最小脚底力 35.7 N、未摔倒，两次失败的 D1 尝试保留。D2 witness: all 6 balls judged correctly, 3 legal balls returned and 3 dead balls correctly let go; physics checks pass. D1 and D3 are also all correct (18 balls over three tiers); in D1, dead balls aimed at the head and torso are avoided by waist lean with a minimum foot force of 35.7 N and no fall, and two failed D1 attempts are retained.

4×

4 倍速 · 原始 18 s，播放 4.5 s4x speed · original 18 s, plays in 4.5 s

TT-D-AA3 死球判断与躯干避让A3 dead-ball judgment with body avoidance

A3执行任务Solve the task

条件Conditions: D2：一条连续运行发出 6 颗独立球（3 个合法来球、3 个擦网死球），沿用为 G1 设计的发球序列；每球周期 3.0 s，共 18 s 仿真；A3 Ultra 不焊接站立；特权状态、无旋。D2: one continuous run with 6 separate balls (3 legal feeds, 3 net-clip dead balls) reusing the launch sequence designed for G1; 3.0 s per ball, 18 s of simulation in total; unwelded A3 Ultra; privileged state, no spin.
任务设计Task design: 同 TT-D-G：合法球回击，死球不得触碰并保持准备姿态；A3 身体更大，避让约束与 G1 不同。场景、发球序列、裁判与验收阈值均未因 A3 修改。As TT-D-G: return legal balls and leave dead balls untouched while holding the ready pose; A3's larger body changes the avoidance constraints. Scene, launch sequence, referee and acceptance thresholds were not changed for A3.
工具Tools: 腕部固连球拍Wrist-mounted racket发球机（6 球弹匣）Ball launcher (6-ball magazine)
Embodiment: A3 Ultra 人形A3 Ultra humanoid
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：18 s 内 6 球全部判对，3 个合法球回击各 100 分，3 个擦网死球靠保持准备姿态正确放过，物理检查通过；D3 也有见证。D1 仍失败（有证据）：3 个合法球回击，2 个死球碰到身体、1 个正确放过；当前控制器与有限搜索未找到安全避让动作，未测试迈步，不构成物理上无法完成的证明。D2 witness: all 6 balls judged correctly within 18 s, the 3 legal balls returned at 100 each and the 3 net-clip dead balls let go by holding the ready pose; physics checks pass, and D3 is also witnessed. D1 still fails with evidence: 3 legal balls returned, 2 dead balls hit the body and 1 is let go correctly; the current controller and a finite search found no safe avoidance motion, stepping was not tested, and this is not a proof of physical impossibility.

该配置的见证视频未随归档保存，暂无公开视频。The witness video of this configuration is not in the archive, so no public video is shown.

TT-R-GH-refG1 五指手持拍连续对打G1 hand-held racket rally

Unitree G1执行任务Solve the task

条件Conditions: D2：N=5，参考对手左右交替回球；G1 右手换为 Inspire 五指手，球拍为独立刚体（拍面同焊接版本，含握柄共 0.175 kg），只靠手指接触握持；开局拍柄平放在朝上张开的掌心，手指闭合、抬起并移到准备姿态后约 8.3 s 发球；特权状态、无旋。D2: N=5 with the reference opponent alternating sides; G1's right hand is an Inspire five-finger hand and the racket is a free rigid body (same blade as the welded version, 0.175 kg with handle) held only by finger contact; the handle starts lying in the open, upturned palm, and the feed fires at about 8.3 s after the fingers close and the racket is lifted to the ready pose; privileged state, no spin.
任务设计Task design: 同 TT-R-G-ref，但球拍不焊接：每次击球冲量对拍柄产生扭矩，握持靠真实接触，拍子滑动或脱手即物理失败。报告相对焊接拍的 C 降幅以及拍柄相对手掌的位移与转角。As TT-R-G-ref, but the racket is not welded: every impact torques the handle, the grip relies on real contact, and a slipping or dropped racket is a physical failure. The drop in C relative to the welded racket and the handle's displacement and rotation in the palm are reported.
工具Tools: 手持球拍（独立刚体，不焊接）Hand-held racket (free rigid body, not welded)参考对手（龙门球拍）Reference opponent (gantry paddle)发球机与拍架Ball launcher and racket rest
Embodiment: G1＋Inspire 五指手G1 with Inspire five-finger hand
是否拒答Refusal: 否No
见证结果Witness result: D2 见证：100 分（C=5/5、H=5、R_end=1），发球后拍柄最大漂移约 1.67 mm，左侧来球使用腰部辅助击球。D1、D3 也为 100 分（D3 拍柄最大漂移 2.27 mm／5.31°），相对焊接拍 C 降幅为 0。D2 witness: 100 (C=5/5, H=5, R_end=1), with a maximum handle drift of about 1.67 mm after the feed and waist-assisted strikes on left-side balls. D1 and D3 also score 100 (D3 maximum handle drift 2.27 mm / 5.31 degrees), and the C drop relative to the welded racket is 0.