https://haiguangboy.com/ 从论文看机器人走向真实世界的下一步 Thu, 03 Sep 2026 04:29:49 GMT https://haiguangboy.com/posts/what_matters_for_latent_actions_in_robot_learning https://haiguangboy.com/posts/what_matters_for_latent_actions_in_robot_learning 潜在动作:光流是负资产 Tue, 01 Sep 2026 00:12:56 GMT https://haiguangboy.com/posts/skild_blog_introducing_s1_in_context_learning_20260817 https://haiguangboy.com/posts/skild_blog_introducing_s1_in_context_learning_20260817 S1:机器人的GPT-3涌现时刻 Wed, 26 Aug 2026 14:25:45 GMT https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation PointWorld:点流统一状态动作 Tue, 25 Aug 2026 23:55:46 GMT https://haiguangboy.com/posts/the_missing_touch_spatially_distributed_tactile_feedback_brings_teleoperation_cl https://haiguangboy.com/posts/the_missing_touch_spatially_distributed_tactile_feedback_brings_teleoperation_cl 指尖麻醉类比:遥操作缺的那味触觉 Tue, 25 Aug 2026 10:41:14 GMT https://haiguangboy.com/posts/flex_pi_a_multi_stream_world_action_model_with_compute_flexibility https://haiguangboy.com/posts/flex_pi_a_multi_stream_world_action_model_with_compute_flexibility FLEX-π:更全面的联合表征 Tue, 25 Aug 2026 01:01:40 GMT https://haiguangboy.com/posts/prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim https://haiguangboy.com/posts/prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim PRISM:规模能否替代显式建模? Mon, 24 Aug 2026 06:24:27 GMT https://haiguangboy.com/posts/hydra-0-action-flow https://haiguangboy.com/posts/hydra-0-action-flow 英伟达把机器人动作定义成"动作流"——机械臂、人手、夹爪全部通用，误差直接降了90.40% Sat, 22 Aug 2026 03:34:36 GMT https://haiguangboy.com/posts/gen-1-5-one-shot https://haiguangboy.com/posts/gen-1-5-one-shot Generalist AI说GEN-1.5不用任何专门设计就学会单样本操作——官方自己写的原文是：10个任务，方差±10%，技能"比微调更脆弱" Fri, 21 Aug 2026 09:40:24 GMT https://haiguangboy.com/posts/omega-eva https://haiguangboy.com/posts/omega-eva SLIM、JEPA-WAM都说别在推理时干活——这篇消融显示：去掉那一步，精化器反而比不精化还差 Wed, 19 Aug 2026 07:29:31 GMT https://haiguangboy.com/posts/slim-05b https://haiguangboy.com/posts/slim-05b 世界模型会预测未来，不等于会控制——给动作能预测后果，给后果能反推动作，只用 0.47B Fri, 14 Aug 2026 00:20:39 GMT https://haiguangboy.com/posts/jepa-wam https://haiguangboy.com/posts/jepa-wam 不靠堆数据 世界模型创新方法 竟让π0.5真机成功率最高提升12.8个百分点 Tue, 11 Aug 2026 07:41:52 GMT https://haiguangboy.com/posts/splatting-physical-scenes https://haiguangboy.com/posts/splatting-physical-scenes 让机器人拿自己当尺子，重建误差 3.31mm，生成式模型差了近 5 倍 Mon, 10 Aug 2026 08:07:47 GMT https://haiguangboy.com/posts/video2robo https://haiguangboy.com/posts/video2robo 一段手机视频合成的数据，真机 62.5%，反超真机采集的 46.7% Sun, 09 Aug 2026 14:42:47 GMT https://haiguangboy.com/posts/joyai-ra-dual-alignment https://haiguangboy.com/posts/joyai-ra-dual-alignment 异构数据的问题不是量，是监督格式不统一 Sun, 09 Aug 2026 05:29:52 GMT https://haiguangboy.com/posts/omega-0 https://haiguangboy.com/posts/omega-0 未来预测越准，机器人反而越迟滞 Sun, 09 Aug 2026 00:43:52 GMT https://haiguangboy.com/posts/transformer-transformer-codesign https://haiguangboy.com/posts/transformer-transformer-codesign 机器人失败，可能不是策略不够大 Sat, 08 Aug 2026 04:56:40 GMT https://haiguangboy.com/posts/rl-100 https://haiguangboy.com/posts/rl-100 RL 不该从零学，该做后训练 Fri, 07 Aug 2026 08:03:15 GMT https://haiguangboy.com/posts/zprl https://haiguangboy.com/posts/zprl RL 该在哪一层介入 Thu, 06 Aug 2026 23:52:06 GMT https://haiguangboy.com/posts/dexora https://haiguangboy.com/posts/dexora 遥操作数据的噪声得进损失函数 Wed, 05 Aug 2026 10:19:52 GMT https://haiguangboy.com/posts/mt3-thousand-tasks https://haiguangboy.com/posts/mt3-thousand-tasks 1 天 1000 任务，靠的是归纳偏置 Wed, 05 Aug 2026 02:54:28 GMT https://haiguangboy.com/posts/pegasus-editable-experience https://haiguangboy.com/posts/pegasus-editable-experience 跨本体迁移该在经验层做 Tue, 04 Aug 2026 08:45:50 GMT https://haiguangboy.com/posts/tacwam https://haiguangboy.com/posts/tacwam 触觉预测只该用来训练，不该给动作看 Mon, 03 Aug 2026 10:18:19 GMT https://haiguangboy.com/posts/archon-whole-body-intelligence https://haiguangboy.com/posts/archon-whole-body-intelligence 李弘扬老师的全身智能高于GR的全身控制 Sat, 01 Aug 2026 22:51:40 GMT https://haiguangboy.com/posts/dr-lfd https://haiguangboy.com/posts/dr-lfd 切换逻辑不该让策略去学 Sat, 01 Aug 2026 15:08:28 GMT https://haiguangboy.com/posts/gemini-robotics-2 https://haiguangboy.com/posts/gemini-robotics-2 被忽略的非共识是全身统一控制 Fri, 31 Jul 2026 04:45:53 GMT https://haiguangboy.com/posts/leapbot-wa https://haiguangboy.com/posts/leapbot-wa 预测式特征不能直接喂扩散模型 Thu, 30 Jul 2026 22:32:33 GMT https://haiguangboy.com/posts/tactile-glove-pose-artifacts https://haiguangboy.com/posts/tactile-glove-pose-artifacts 手套触觉读数里混着手的姿态 Wed, 29 Jul 2026 14:27:41 GMT https://haiguangboy.com/posts/cd-lam https://haiguangboy.com/posts/cd-lam 世界模型不听话，是潜在动作被污染了 Wed, 29 Jul 2026 12:37:27 GMT https://haiguangboy.com/posts/pi-rl-chelsea-finn-talk https://haiguangboy.com/posts/pi-rl-chelsea-finn-talk 机器人 RL 的瓶颈是采样不是算法 Wed, 29 Jul 2026 08:59:36 GMT https://haiguangboy.com/posts/n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens https://haiguangboy.com/posts/n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens 触觉当预测目标而非观测输入 Wed, 29 Jul 2026 00:55:47 GMT https://haiguangboy.com/posts/robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode https://haiguangboy.com/posts/robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode 不是"要不要世界模型"，是"它吐出来的东西有没有结构" Fri, 24 Jul 2026 03:40:57 GMT https://haiguangboy.com/posts/liangwenfeng-world-model https://haiguangboy.com/posts/liangwenfeng-world-model 这句话缺了一个前提：你的数据管道通没通 Thu, 23 Jul 2026 07:00:13 GMT https://haiguangboy.com/posts/real_time_semg_based_telecontrol_of_an_assistive_robotic_arm_using_a_1d_convolut https://haiguangboy.com/posts/real_time_semg_based_telecontrol_of_an_assistive_robotic_arm_using_a_1d_convolut 神经信号：它的价值和它的脆弱，来自同一处 Thu, 23 Jul 2026 02:54:07 GMT https://haiguangboy.com/posts/sudu-third-position https://haiguangboy.com/posts/sudu-third-position 苏度科技：横切一刀的第三种位置 Thu, 23 Jul 2026 01:17:33 GMT https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling LAWM：为什么动作标签会成为负担 Wed, 22 Jul 2026 02:41:01 GMT https://haiguangboy.com/posts/qianxun-general-brain-data-loop https://haiguangboy.com/posts/qianxun-general-brain-data-loop 千寻智能的数据闭环赌注 Tue, 21 Jul 2026 14:35:35 GMT https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation T-Rex：触觉为何要单独建模 Tue, 21 Jul 2026 09:41:47 GMT https://haiguangboy.com/posts/lingchu-optical https://haiguangboy.com/posts/lingchu-optical 灵初智能：数据策略要看场景复杂度 Sat, 18 Jul 2026 11:32:44 GMT https://haiguangboy.com/posts/sunday-act2 https://haiguangboy.com/posts/sunday-act2 Sunday的资源集中战略 Sat, 18 Jul 2026 03:28:45 GMT https://haiguangboy.com/posts/xsquare-robot https://haiguangboy.com/posts/xsquare-robot 自变量：信念趋同不是证据趋同 Sat, 18 Jul 2026 01:00:29 GMT https://haiguangboy.com/posts/xuhuazhe-poke https://haiguangboy.com/posts/xuhuazhe-poke 许华哲的组织下注 Sat, 18 Jul 2026 00:59:52 GMT https://haiguangboy.com/posts/schrodinger-navigator https://haiguangboy.com/posts/schrodinger-navigator 零样本导航：Schrödinger's Navigator Sat, 18 Jul 2026 00:59:17 GMT https://haiguangboy.com/posts/psi-zero https://haiguangboy.com/posts/psi-zero 少量机器人数据下的策略学习 Tue, 07 Jul 2026 11:22:32 GMT https://haiguangboy.com/posts/fast-wam https://haiguangboy.com/posts/fast-wam 训练视频目标，比测试时想象未来更关键 Tue, 07 Jul 2026 07:32:00 GMT https://haiguangboy.com/posts/orca https://haiguangboy.com/posts/orca 世界模型的关键是可读出的状态 Mon, 06 Jul 2026 15:19:24 GMT