LLM-based robotic agent systems can generate programs for robot control, yet their generalization to new tasks often depends on elaborate scaffolding. Expanding their skill repertoires alone offers limited guidance on how to assess new capabilities or identify what remains beyond their reach. We introduce META-Robot, a framework that couples self cognition with capability scaling by expanding the criteria through which an agent evaluates and develops its skills. At its core, action-level predicates make execution outcomes explicit and available for LLM reasoning. The agent uses these assessments to identify its own limitations, extend relevant evaluation criteria, and guide the structured expansion of its skills. Proprioceptive feedback further grounds this process in the robot's physical state, enriching the evidence available for self-cognition. Experiments in simulation and on real robots demonstrate stronger overall task performance, generalization to new tasks, continued expansion of skills and evaluation criteria, and effective use of non-visual proprioceptive signals.
Figures & tables
Figure 1: META-Robot starts from a task description. Its framework consists of three components: (1) Loop Execution Agent, which generates programs and predicates, evaluates execution evidence, and repairs failures; (2) Proprioceptive Signal Orchestration, which supplements insufficient visual observations with physical feedback; and (3) coupled Predicate and Skill Libraries, which consolidate validated criteria and repairs for cognitive and capability scaling.
Figure 2: META-Robot Loop Execution Agent. Given a task description, the agent (1) decomposes it into executable subtasks and initial conditions; (2) jointly generates the task program and action-level predicates; (3) executes each skill while collecting task-relevant evidence; (4) evaluates predicate transitions to determine whether execution should continue or be repaired; (5) locally corrects and re-executes the failed skill, missing evidence routine, or verification check; and (6) performs system-level repair and full-task validation, consolidating validated predicates and skills for reuse. This procedure is repeated to form a closed execution–evaluation–repair loop.
Figure 3: Proprioceptive Signal Orchestration Mechanism Example: in thin-towel grasping, torque change provides a contact cue, while gripper feedback and a lift check verify successful acquisition to complement insufficient visual evidence during execution and verification. Validated signal usage is then consolidated into a reusable skill together with corresponding action predicates, jointly expanding sensing, execution, and verification capabilities
Category
Task Type
OpenVLA
π0
π0.5
CaP- Agent0
ASPIRE
Playful / RATS
META-Robot (DeepSeek- V4-Flash)
META-Robot (GPT-6 Astra)
Object
Pos
0.00
0.00
17.00
22.00
98.00
61.00
98.00
99.60
Task
0.00
0.00
1.00
18.00
95.00
63.00
88.60
99.00
Goal
Pos
0.00
0.00
38.00
26.00
81.00
43.00
72.20
85.80
Task
0.00
0.00
0.00
17.00
45.00
36.00
64.60
69.80
Spatial
Pos
0.00
0.00
20.00
12.00
51.00
29.00
83.00
89.25
Task
0.00
0.00
1.00
14.00
60.00
31.00
88.40
95.80
Table 1: LIBERO-Pro evaluates the success rate of each model across Object, Goal, and Spatial tasks
Figure 4: (a) illustrates the robot performing diverse real-world tasks, including precise insertion, pick-and-place, proprioceptive signals task and long-horizon instruction following. (b) shows the success rate of these tasks.
Figure 5: Zero-shot transfer to LIBERO-Pro Long under different library retention ratios. Success rates are averaged over ten tasks.
Category
Task Type
Full META-Robot (%)
Without Predicate (%)
Object
Task
99.00
85.00
Pos
99.60
91.00
Goal
Task
69.80
56.00
Pos
85.80
69.00
Spatial
Task
95.80
76.00
Pos
89.25
72.00
Table 2: Ablation Study of META-Robot across Different Tasks on LIBERO-Pro: We compare the task success rates (%) of full META-Robot and its variant without action-level predicates across the Object, Goal, and Spatial suites under task and position perturbations. Removing predicates consistently reduces performance across all six settings, demonstrating their contribution to execution assessment and targeted repair.
Figure 6: Effect of Proprioceptive Feedback on Manipulation Success: (a) shows the execution process during the simulation, while (b) displays the success rates before and after the ablation of the agent’s signals.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: We validate the pick-and-place ability of META-Robot on a real-world Galaxea R1 Lite. A yellow tennis ball and an open cardboard box sit on the table; only the right arm moves. The operator names neither a grasp nor a drop point. We find META-Robot can easily complete this task. We find that META-Robot can pick and place from two live object poses rather than from a saved trajectory or large training data.
Figure 8: We validate META-Robot’s long-horizon manipulation capability on a real-world Galaxea R1 Lite. A peach, a tennis ball, an apple, and an open cardboard box sit on the table; only the right arm moves. This task shows the robot’s ability to perform long-horizon pick-and-place tasks.
Figure 9: We validate the order-following ability of Meta-Robot on a real-world Galaxea R1 Lite using three inverted plastic cups of the same size—orange, purple, and green. The focus of this task is the robot’s ability to follow the specified order, rather than the individual grasping and transfer motions. This real-world task evaluates generalization across different specified color orders.
Figure 10: We validate the ability of META-Robot to fold a soft object on a real-world Galaxea R1 Lite using one flat purple cloth on the table. This task requires the use of proprioceptive signals as the towel is very thin and is quite difficult to grasp using visual feedback alone. META-Robot gives the plan: a bimanual first fold that turns the square into a half-width rectangle, then a right-arm second fold that turns that rectangle into a quarter. Both arms move in the first half; only the right arm moves in the second. The first episode makes the half-fold; the second episode makes the quarter. We validate its ability to handle complex tasks.
Figure 11: We validate the breakfast task of META-Robot on a real-world Galaxea R1 Lite. A brown toast slice, a small black Oreo, and a pretzel sit on the table; only one arm moves. We use a fixed two-layer order: Oreo onto the bread, then pretzel on top of that stack. This task is also an evaluation of the system’s use of proprioceptive feedback.
Figure 12: We validate the weighing capability of META-Robot on a real-world Galaxea R1 Lite using three visually identical plastic bottles whose fill levels—empty, half, and full—cannot be recovered from RGB appearance. META-Robot uses its proprioceptive signal to realize this task with high accuracy.
Figure 13: We validate a precision plug-insertion task of META-Robot on a real-world Galaxea R1 Lite. A small three-pin plug and a white power strip are placed at randomized poses on the table. This task shows the use of proprioceptive signal in precise task.
Figure 14: Zero-shot transfer to LIBERO-Pro Long tasks: we evaluate ten representative manipulation tasks under 25%, 50%, 75%, and 100% information retention. The tasks include T01 soup-and-sauce placement, T02 cheese-and-butter placement, T03 stove/moka-pot interaction, T04 bowl-in-drawer placement, T05 mug-to-plate placement, T06 book-to-caddy placement, T07 mug-and-pudding arrangement, T08 soup-and-cheese placement, T09 moka-pot placement, and T10 mug-in-microwave placement. Each radar plot reports per-task success rates, with the mean success rate shown below.
Figure 15: We extend the base model comparison shown in CaP-X Fu et al. [2026] with DeepSeek-V4-Flash and GPT-6 Astra, reporting task success rates of 31% and 45%, respectively.
Multi-agent large language model (LLM) systems have shown promise for solving complex tasks through agent collaboration. However, existing frameworks assign tasks based on predefined roles without considering whether an agent can accurately assess its own competence boundaries, leading to overconfident execution on tasks beyond its expertise. Inspired by metacognition theory from cognitive science, we propose MetaCogAgent, a multi-agent LLM framework where each agent is equipped with a Metacognitive Self-Assessment Unit that evaluates task-capability alignment before execution. The framework introduces three contributions: (1) a self-assessment mechanism that estimates per-task confidence by combining verbalized uncertainty with historical capability profiles; (2) an adaptive delegation protocol that routes low-confidence tasks to better-suited agents through cross-agent evaluation; and (3) a capability boundary learning module that iteratively refines each agent's competence model via cybernetic feedback. Experiments on our constructed MetaCog-Eval benchmark (700 tasks across 5 cognitive dimensions) demonstrate that MetaCogAgent achieves 82.4% task accuracy -- 8.7% above the best routing baseline -- while using 5% fewer API calls than AutoGen and 34% fewer than ensemble voting. Ablation studies confirm that each metacognitive component contributes to overall system performance.
Chenyu Wang, Yang Shu
School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou, China · Zhejiang University, Hangzhou, China
Agentic AI based on Large Language Model generalization capabilities offers a wide range of potential applications, including planning for embodied tasks. For example, embodied agents based on Foundation models can generate plausible plans in autonomous robotics scenarios. Due to limited context windows or hallucinatory phenomena in the next-token prediction formulation, behaviors may be generated without establishing whether the deployed robot and the observed environment actually support the requested operation, in what we call a "grounding failure". Thanks to the recent improvements in reasoning capabilities of foundation models, autonomous robot behavior generation problem can be formulated as a code generation problem. We present iAm.md, a Markdown standard and generation framework, that allows anchoring this process in complementary forms of deployment evidence. Through open-vocabulary semantic mapping, we combine local vision-language detections and object segmentation and refer them to persistent object records in this intermediate standardized representation, allowing agentic introspection. We then study this new technique on a simulated TIAGo, on navigation-and-manipulation tasks, showing how this standardized representation jointly supports skill self-assessment and executable task generalization.
Building reliable robot capabilities across diverse tasks requires substantial human effort to develop and maintain skills, design rewards, and integrate perception with control. We present Reconstruct, Practice, Go Real (RPG), a framework for autonomous improvement of robot execution systems without updating model weights. RPG identifies manipulation capabilities in an offline dataset and constructs related practice tasks in simulation. During practice, RPG uses execution feedback, privileged simulator state, and available dataset videos to diagnose failures. It develops new reusable symbolic skills, refines existing skills, and revises the system prompt based on these diagnoses. Cross-task evaluation tests individual candidate changes and merged revisions before they are retained for reuse. At test time, a multimodal LLM uses the resulting system prompt and skill library to coordinate perception and robot control. On held-out initializations of 22 manipulation tasks, RPG improves task success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming all evaluated baselines, including ASPIRE (75.5%) and CaP-Agent0 powered by GPT-6 Astra Pro (60.0%). After a common calibration and hardware-adaptation procedure, the frozen system succeeds in all 30 physical trials, with ten trials on each of three tasks. Project Website: https://rpg-robot.github.io/