The 'Double Helix' of Embodied Intelligence, Happening in Space?
具身智能的“双螺旋”，发生在空间里？
The StarSource Intelligence team, in collaboration with universities and research institutions, has launched two studies, DA-Nav and AdaOcc, addressing the compatibility issues between robot navigation and three-dimensional perception. DA-Nav achieves city-level navigation through discrete image point selection and chain-of-thought reasoning, while AdaOcc enhances the flexibility and robustness of three-dimensional perception through adaptive geometric guidance and sparse point representation. These two works demonstrate key technical pathways for embodied intelligence in real-world environments.
Source: RoboSpeak — WeChat · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
59%
View original evidence
在239条长距离闭环轨迹的测试中，DA-Nav的导航成功率达到59.00%，偏差出现后的纠偏成功率高达98.15%，相比之下，其他对比方案的纠偏成功率普遍只有个位数。
Open source S20Reported success rate
98.15%
View original evidence
在239条长距离闭环轨迹的测试中，DA-Nav的导航成功率达到59.00%，偏差出现后的纠偏成功率高达98.15%，相比之下，其他对比方案的纠偏成功率普遍只有个位数。
Open source S20navigation distance
>1,200
View original evidence
团队把训练好的模型直接部署在厂商四足机器人和人形机器人上，没有在真实数据上做任何微调，就在不同城市环境中完成了超过1200米的稳定导航。
Open source S21number of trajectories
239
Trajectory count
View original evidence
在239条长距离闭环轨迹的测试中，DA-Nav的导航成功率达到59.00%，偏差出现后的纠偏成功率高达98.15%，相比之下，其他对比方案的纠偏成功率普遍只有个位数。
Open source S20number of sequences
≈286,000
View original evidence
整个数据集包含约28.6万序列样本，其中约12.8万帧是恢复行为数据，与15.8万帧的正常导航数据混合训练后，模型学会了面对偏离时如何拉回正轨。
Open source S19number of frames
≈128,000 frames
View original evidence
整个数据集包含约28.6万序列样本，其中约12.8万帧是恢复行为数据，与15.8万帧的正常导航数据混合训练后，模型学会了面对偏离时如何拉回正轨。
Open source S19number of frames
158,000 frames
View original evidence
整个数据集包含约28.6万序列样本，其中约12.8万帧是恢复行为数据，与15.8万帧的正常导航数据混合训练后，模型学会了面对偏离时如何拉回正轨。
Open source S19
- dataset: not_reported
- code: not_reported
- weights: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
在239条长距离闭环轨迹的测试中，DA-Nav的导航成功率达到59.00%，偏差出现后的纠偏成功率高达98.15%，相比之下，其他对比方案的纠偏成功率普遍只有个位数。
Open source S20
团队把训练好的模型直接部署在厂商四足机器人和人形机器人上，没有在真实数据上做任何微调，就在不同城市环境中完成了超过1200米的稳定导航。
Open source S21
整个数据集包含约28.6万序列样本，其中约12.8万帧是恢复行为数据，与15.8万帧的正常导航数据混合训练后，模型学会了面对偏离时如何拉回正轨。
Open source S19
Source：RoboSpeak — WeChat · mp.weixin.qq.com