Repository navigation
Replies: 3 comments
|
基于这里的讨论，已完成第一个可运行任务：豆腐状软块对半切，按用户选择只做刀具—材料物理闭环及切后分离。实现依据 BladeMaster v2，代码已合入 PR #195，原始记录和视频在 v0.61.0。 默认结果是两块质量各占 50%；撤刀后对一侧施加已知诊断冲量，相对运动 3.221 mm。半切与未切负对照都仍连通。半时间步和细网格工况也能分离，但反力峰值对时间步敏感，尚不能证明切割力收敛或真实豆腐准确性。全部 8 次尝试（含失败和修正）均保留。 这为最初的问题增加了一项实际证据：仿真现在可以在这个合成单次切割任务里表达“切开后不再重新连接”，并检查 1/2 质量比例。它还没有验证 1/4、多刀、香蕉、机械臂操作或真实材料参数。真实实验的下一价值是识别材料与切割阻力参数，再独立对照切割力和质量比例；不是用本次动画替代真机证据。 中文方法、结果与复现 · English · 完整验收 #192 English: The first paper-based CPU task is delivered: synthetic tofu-like block halving with two-way dynamic blade/material coupling and persistent separation. The default yields 50/50 mass and 3.221 mm relative motion under a known one-sided post-withdrawal impulse. Partial/uncut controls remain connected; timestep/grid sensitivity cases also separate. Peak forces are timestep-sensitive, so no force convergence, calibrated real-food accuracy, quarter-cut accuracy, GPU speed or robot-policy claim follows. All eight attempts and independent saved-record scoring are public. This updates the earlier discussion with measured evidence while preserving its historical proposals. |
|
豆腐与香蕉切割展示已更新：v0.62.0 · PR #201。
版本附带全部 25 次切割运行及失败记录、源码快照、独立评分、视频和 SHA-256。完整仓库门槛及双语站部署通过。当前仍是合成材料、刀具—材料闭环，不含机器人策略或真实食材标定。三列任务已对齐，但原作完整参数和 gap/volume 指标尚未复现；后续 #200 单独跟踪，因此不作同条件优劣排名。 English: The comparison now uses the same two-cut tofu task in all three columns. The original two videos remain attributed to the BladeMaster project; DexLab provides a separate recorded run. Banana slicing is its own three-cut/four-piece experiment, not a quarter-mass result. Raw evidence, failures, source snapshots and checksums are available in v0.62.0. Synthetic parameters and authored surfaces do not establish real-food or robot accuracy; complete controlled-comparison inputs remain open in #200. |
|
切割展示精修已交付：#203 / #207 / v0.64.0。 README 新增逐帧数值 GIF；三列仍统一为豆腐两刀，香蕉独立展示。切口、内部体积和切后位移取自保存状态，作者视频和作者报告数字保留来源，未当作本项目实测值。修正了香蕉几何平滑造成的圆钝切面和跨块表面连接；新细采样记录、半步长及负对照、失败导出和原始数据均已保留。 这轮诊断说明，表面几何、材料响应和独立运动必须分别检查：好看的分离画面不能代替原生连接图与动量证据。564 个表面帧与原生连通数一致，303 个导出帧的数值经独立计算核验；完整仓库验收、线上版本及 8 个发布资产已读回确认。 效果对比 · 同状态修正前后与论文细节。作者同条件设置缺口仍由 #200 跟踪；论文香蕉塑性、材料／力值和完整重建仍由 #204 跟踪。这轮未完成真实食材标定或完整 BladeMaster 复现。 |
Uh oh!
There was an error while loading. Please reload this page.
初步判断：切割仿真已经存在；如果目标是固定食材上稳定切出 1/2、1/4，可以先做真机基线。需要跨材料泛化、优化受力或大量策略试验时，再用经过实测校准的切割仿真。
本帖讨论一个具体问题：机器人能否把东西按指定比例切分，DexLab 应该怎样帮助选择合适的仿真和验证方法？先按食材／软材料理解；金属、木材加工需要另外的模型与工艺。
2026-10-14，由维护者授权的 AI 助手整理公开论文、作者代码和 DexLab 当前状态。以下区分作者报告、仓库已交付能力和建议；没有运行或独立复现这些切割系统，建议尚未采纳为开发排期。
1. “切 1/4”首先要定义清楚
“取一块 1/4”和“四等分”是不同任务。形状与密度满足相应对称条件时，中心切一刀、再正交切一刀可以得到等体积／等质量的四份；一般物体没有这个保证。均匀材料可以按累计体积寻找切面；不应把包围盒长度的 25% 当成质量的 25%。
2. 目前有哪些可参考的仿真？
下面是针对这个问题的候选，不是完整排名。支持形变、SDF 碰撞或 MPM，并不自动证明支持切断、切面接触和切后独立操作。
sim与物理taichi：前者估计切出质量／碰撞，不能估计切割能耗。环境说明另外，RoboChop 展示了用视觉跟踪切后物体、组合动作原语完成不同数量／类型切片的真机路线。这支持“先把任务闭环做起来”的思路，但其摘要中的切片规格不能作为等质量精度证据。作者论文摘要
这里需要分开三层：几何切分代理 → 有材料响应的切割过程 → 完整机器人闭环。把模型替换成预切好的两块，可以用于流程验证；如果要评价刀具受力、断裂、卡刀或材料泛化，就需要额外的物理模型与实测证据。
3. 能不能直接用真机解决？
可以。我的建议是：任务目标清楚、材料和刀具固定时，先用受控真机基线确定需求与主要误差来源。 这是工程判断，不是本帖已经完成的实验。
一条最小路线是：
仿真适合承担三种增量价值：减少耗材和试切次数；筛选速度／角度／横向运动等策略；研究材料、刀具、固定方式变化后的失败。可以先记录少量真机数据，再选最简单且足够预测目标指标的模型，形成 real → sim → real；高保真切割无需成为第一个真机演示的前置条件。
反过来，直接在真机上大量从零探索，也未必更省成本：每次切割改变样本，重置、耗材和跨批次材料变化都要计入。仿真中拟合一条力曲线也不代表泛化，应保留新物体、速度或材料批次用于独立检验。
4. 放到 DexLab 中，具体应该研究什么？
核对 main@7e4a077 的 README、文件树和当前工作入口，目前未见已交付的切割任务及 1/2、1/4 比例验收。现有接触、摩擦、SDF 和布料证据不能替代切割资格；#6 的真机标定也仍有实测输入缺口。
建议沿用 #189 的“任务 → 起步策略 → 仿真配置 → 物理验证”，把问题收敛为：
可以比较三条有相同任务定义的路线：几何规划＋真机闭环、经实测标定的简化力模型、专用切割物理。分别记录模型准备／标定成本、运行成本、真实试切次数和保留测试集的预测能力。不要只比较视频效果或模拟内部奖励，也不要预设复杂模型一定获胜。
如果需要机器人仿真与材料仿真耦合，必须检查刀具运动传入、材料反力返回、坐标／时钟及避免重复计算接触。仅同步两边画面不构成动力学耦合。具体后端接入另行设计；不把文献系统算作已经接入的 UniSim 或 DexLab 能力。
建议的最小验收维度：
m0，目标份额r，分离后对应块质量m；误差abs(m/m0 - r)。四等分报告全部四块，并单列损失；体积目标另用体积评分容差应由任务需求和称重／几何测量不确定性确定，先冻结再评测，本帖不虚设一个“通用合格率”。若未来检查现有六引擎，分别记录支持、未验证和不适用；缺少切割模型不能记成“机器人切割任务失败”，新系统也不自动成为第七个已准入引擎。
5. 希望大家补充的三个问题
本帖交付的是可讨论的选型与验证建议。开发顺序继续以当前工作入口为准；本帖未启动实验、硬件运动或后台调度。
English discussion
Cutting simulation exists. For a fixed material and tool, a hardware baseline can be a practical first step toward halves or quarters; calibrated simulation becomes useful when it reduces physical trials or predicts force, failure and transfer across conditions. This is an engineering recommendation, not a reproduced result.
Prepared by an AI assistant at the maintainer’s request on 2026-10-14, using public papers, author repositories and current DexLab records. No cutting system was executed or independently reproduced. The working assumption is food or soft materials; metal and wood machining require different models and processes.
Define the fraction first. A quarter of the length, volume and mass are different objectives. Removing one quarter also differs from producing four equal pieces. Image area or bounding-box length does not generally determine mass. Uniform density permits volume-based planning; heterogeneous density, cores, cavities and material loss require additional treatment and independent measurement.
Relevant evidence:
Suggested hardware starting point: fix a relatively uniform material and a stable fixture; define volume versus mass and one portion versus all portions; measure initial mass and geometry; plan a simple cut; execute with qualified compliance/contact handling; verify complete separation and weigh every required piece. Reobserve before successive cuts. Post-cut weighing improves evaluation and subsequent decisions but cannot undo an undersized piece. Treat corrective trimming as a separately declared task condition.
The DexLab question should be about necessary fidelity: which model is sufficient to predict actual portion error, forces and failures, and how many physical trials does it save? Compare geometry plus hardware feedback, a calibrated reduced force model, and specialized cutting physics using the same task definition and held-out conditions. Include model preparation, calibration and runtime costs. A more elaborate simulator need not win.
At main@7e4a077, the inspected README, tree and current-work entry did not show a delivered fractional-cutting task. Existing contact/SDF/cloth results are not cutting qualification. Follow Discussion #189: task → starting strategy → simulation configuration → physical validation.
Measure each target portion using
abs(m/m0 - r)for mass fraction, report all pieces for quartering and account for loss separately. Check complete separation, independent piece motion, measured force/displacement, slip, jams, incomplete cuts, time and sensitivity to discretization. A volume target needs its own volume metric. Set tolerances from the application and measurement uncertainty before evaluation; keep calibration and held-out data separate. If robot and material simulators are coupled, verify motion/wrench exchange, timing, frames and absence of duplicate contact forces.Open questions: What material and fraction definition are intended? Is the priority a fixed-scene hardware demonstration or generalization across materials/tools? Are there shareable trajectories, force logs, before/after masses or reproduction experience?
This is a discussion proposal. It does not change the current work order or start development, simulation runs, hardware motion or scheduling.
All reactions