Notable Research at CoRL 2026 - The Evolution of Robot AI through VLA, World Models, and LLM Agents
Notable Research at CoRL 2026 - The Evolution of Robot AI through VLA, World Models, and LLM Agents
This article summarizes notable research presented at CoRL 2026, an international conference on robot learning held in Austin, Texas from November 9 to 12, 2026. The content covers key advancements in Vision-Language-Action (VLA) models such as π₀.₇, MolmoAct2, and RoboTTT; world action models including DreamZero, GR00T 2, ImageWAM, LaWAM, and τ₀-WM; LLM and AI agent applications like ENPIRE and MEM; and techniques in imitation learning and reinforcement learning such as EgoScale and SuperMimic. The author acts as a relayer, compiling summaries of official project pages and conference materials without independent verification.
This article summarizes notable research presented at CoRL 2026, covering Vision-Language-Action (VLA) models, world action models, and LLM-based agents; claims regarding specific performance metrics and capabilities are attributed to the respective project pages or conference proceedings.
Source: Robotics — Robot learning news discovery · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
40%
Reported zero-shot result; the excerpt does not specify an evaluation protocol.
View original evidence
dy movements It is reported that for 50 tasks the target robot had not experienced before, it recorded an average success rate of 40% zero-shot, which improved to 95% with additional learning using 5 demonstration data points per task. Furthermore, in a configuration where an inference model plans the work and GR00T 2 handles object manipulation, it recorded 97-98 successes out of 100 for each of 3 types of toy assem
Open source S12successes out of 100
97–98
View original evidence
dy movements It is reported that for 50 tasks the target robot had not experienced before, it recorded an average success rate of 40% zero-shot, which improved to 95% with additional learning using 5 demonstration data points per task. Furthermore, in a configuration where an inference model plans the work and GR00T 2 handles object manipulation, it recorded 97-98 successes out of 100 for each of 3 types of toy assem
Open source S12
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: 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.
dy movements It is reported that for 50 tasks the target robot had not experienced before, it recorded an average success rate of 40% zero-shot, which improved to 95% with additional learning using 5 demonstration data points per task. Furthermore, in a configuration where an inference model plans the work and GR00T 2 handles object manipulation, it recorded 97-98 successes out of 100 for each of 3 types of toy assem
Open source S12
Source：Robotics — Robot learning news discovery · note.com