Robots Learning to Think: Cutting-Edge Robotics Research

Heng, Tang, Xu, Bao, Huang & Wang (2026). Dexterous manipulation with humanoid hands has required enormous amounts of task-specific demonstration data. HumDex introduces a data-efficient framework combining motion retargeting from human video with in-simulation augmentation, achieving human-level performance on a suite of contact-rich manipulation benchmarks with a fraction of prior training data.

Xie, Qi & Sadigh (2026). Piano playing demands the highest dexterity of any keyboard instrument — each note requires precise finger placement, pressure and timing. HandelBot demonstrates that policies trained in simulation can transfer to real pianos via fast adaptation to the acoustic feedback signal, achieving musically coherent performances across multiple pieces not seen during training.

Liu, Zhou, Chi, Han, Rong, Chen, Wang, Wang & Zhang (2026). Vision-language-action (VLA) models for robotics have been passive — they observe and act but do not actively seek information. SaPaVe introduces active perception into the VLA framework: the robot autonomously decides when and where to look based on task uncertainty, dramatically improving performance on partially observable manipulation tasks.

Borse, Xie, Huang & Jin (2026). Contact simulation is the key bottleneck for training contact-rich robot manipulation policies at scale. ComFree-Sim's analytical (non-iterative) contact model enables GPU-parallel simulation of thousands of contact-rich scenarios simultaneously, providing the data throughput needed for sample-efficient learning of complex manipulation skills.

Duan, Shi, Teng, Zhao, Zhang, Li & Yang (2026). 3D occupancy prediction from 360-degree cameras is a critical capability for mobile robots and autonomous vehicles. O3N extends occupancy prediction to open vocabulary, allowing the system to detect and localise objects outside its training distribution — essential for real-world deployment in environments that don't match controlled datasets.

Lin et al. (2026). Hardware design matters as much as software in dexterous robotics. CRAFT introduces a tendon-driven robotic hand with a hybrid compliant architecture — rigid finger bones with soft distal pads — that provides the precision of rigid linkages with the safety and adaptability of soft robotics. Demonstrated on a range of manipulation tasks including cloth handling.

Liu, Wu, Chi, Cai, Hung, Yu, Li, Hu, Rao & Duan (2026). Test-time adaptation of spatial representations is critical for robots operating in changing environments. Spatial-TTT adapts depth and occupancy representations online from streaming video, maintaining accurate 3D world models even as scenes change — a key capability for household robots that must handle dynamic clutter.

Surynek (2026). Automated 3D printing of multi-object batches requires solving combinatorial planning and scheduling problems at industrial scale. This paper's parallel portfolio of CEGAR strategies provides a practical solution that outperforms single-strategy approaches on real manufacturing test cases, demonstrating how AI planning can be directly integrated into production workflows.

Paul & Regli (2026). While focused on data pipelines rather than physical manipulation, WORKSWORLD addresses the same core challenge as many robotics planning problems: jointly optimising what to do (task planning) with when and where to do it (scheduling) under resource constraints. The framework generalises to physical robot workflow orchestration in multi-robot warehouse settings.

Heng, Tang, Xu, Bao, Huang & Wang (2026). Dexterous manipulation with humanoid hands has required enormous amounts of task-specific demonstration data. HumDex introduces a data-efficient framework combining motion retargeting from human video with in-simulation augmentation, achieving human-level performance on a suite of contact-rich manipulation benchmarks with a fraction of prior training data.

Xie, Qi & Sadigh (2026). Piano playing demands the highest dexterity of any keyboard instrument — each note requires precise finger placement, pressure and timing. HandelBot demonstrates that policies trained in simulation can transfer to real pianos via fast adaptation to the acoustic feedback signal, achieving musically coherent performances across multiple pieces not seen during training.

Liu, Zhou, Chi, Han, Rong, Chen, Wang, Wang & Zhang (2026). Vision-language-action (VLA) models for robotics have been passive — they observe and act but do not actively seek information. SaPaVe introduces active perception into the VLA framework: the robot autonomously decides when and where to look based on task uncertainty, dramatically improving performance on partially observable manipulation tasks.

Borse, Xie, Huang & Jin (2026). Contact simulation is the key bottleneck for training contact-rich robot manipulation policies at scale. ComFree-Sim's analytical (non-iterative) contact model enables GPU-parallel simulation of thousands of contact-rich scenarios simultaneously, providing the data throughput needed for sample-efficient learning of complex manipulation skills.

Duan, Shi, Teng, Zhao, Zhang, Li & Yang (2026). 3D occupancy prediction from 360-degree cameras is a critical capability for mobile robots and autonomous vehicles. O3N extends occupancy prediction to open vocabulary, allowing the system to detect and localise objects outside its training distribution — essential for real-world deployment in environments that don't match controlled datasets.

Lin et al. (2026). Hardware design matters as much as software in dexterous robotics. CRAFT introduces a tendon-driven robotic hand with a hybrid compliant architecture — rigid finger bones with soft distal pads — that provides the precision of rigid linkages with the safety and adaptability of soft robotics. Demonstrated on a range of manipulation tasks including cloth handling.

Liu, Wu, Chi, Cai, Hung, Yu, Li, Hu, Rao & Duan (2026). Test-time adaptation of spatial representations is critical for robots operating in changing environments. Spatial-TTT adapts depth and occupancy representations online from streaming video, maintaining accurate 3D world models even as scenes change — a key capability for household robots that must handle dynamic clutter.

Surynek (2026). Automated 3D printing of multi-object batches requires solving combinatorial planning and scheduling problems at industrial scale. This paper's parallel portfolio of CEGAR strategies provides a practical solution that outperforms single-strategy approaches on real manufacturing test cases, demonstrating how AI planning can be directly integrated into production workflows.

Paul & Regli (2026). While focused on data pipelines rather than physical manipulation, WORKSWORLD addresses the same core challenge as many robotics planning problems: jointly optimising what to do (task planning) with when and where to do it (scheduling) under resource constraints. The framework generalises to physical robot workflow orchestration in multi-robot warehouse settings.