# Robotic Agent Survey: Four Directions in One Article

Robotic Agent Survey: Four Directions in One Article

Full article: https://haiguangboy.com/posts/towards-agentic-robots

## Core Methods

Covers training, perception, planning, control, and adaptation.

The focus is not just predicting actions, but also adjusting plans based on new observations, selecting executable skills, and deciding after failure whether to gather more observations, change the plan, or update the policy.

This part helps distinguish: whether what needs improvement is model capability or the decision-making approach, avoiding the default of retraining whenever failure occurs.

## Key Results

Covers collection, annotation, and refinement.

Agents can schedule supplementary collection based on insufficient coverage or failure stages; failed trajectories do not have to be discarded entirely—HELP and ARS filter out the effective segments within them.

This part connects what to collect, how to annotate, and which segment to keep, explaining how the data loop revolves around actual weaknesses.

## Related Route Comparison

- Same route · [1 day, 1000 tasks, relying on inductive bias](https://haiguangboy.com/posts/mt3-thousand-tasks) `learning_a_thousand_tasks_in_a_day_2026_08`: repair methods should correspond to failure sources validates failure attribution: retrieval 22.3%, pose estimation 23.9%, segmentation 19.5%, grasped object 20.2%
- Same route · [Gemini_Robotics_2_Safety](https://haiguangboy.com/posts/gemini_robotics_2_safety) `gemini_robotics_2_safety_2026_07`: Agentic does not automatically mean a higher level of autonomy validates ASIMOV-Agentic's positioning: only tests high-level semantic orchestration, explicitly does not test functional safety architecture
- Same route · [Introducing Gemini Robotics ER 2](https://haiguangboy.com/posts/google_blog_gemini_robotics_er_2_20260730) `google_blog_gemini_robotics_er_2_20260730_2026_07`: Agentic does not automatically mean a higher level of autonomy validates architectural statement: VLA models and navigation APIs are declared side by side as the agent's tools
- Same route · [PAVE: failure data can also feed the policy](https://haiguangboy.com/posts/pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies) `pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies_2026_09`: PAVE: failure data can also feed the policy
- Same route · [PointWorld: point streams unify state and action](https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation) `pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation_2026_08`: PointWorld: point streams unify state and action
- Same route · [RL should not learn from scratch, it should do post-training](https://haiguangboy.com/posts/rl-100) `rl_100_performant_robotic_manipulation_with_real_world_reinforcement_learning_2026_08`: skill evolution includes promotion, validation, and deprecation mechanisms validates using an approximate model Q-function as a gate during the offline stage, only advancing the behavior policy if the prediction genuinely improves, otherwise rejecting that update

## Boundaries

⚠️ A research map, not unified validation of an optimal architecture
The four functional domains are not four modules that must be deployed independently; the same work can span multiple categories. The cases come from different studies, and their respective results cannot be pieced together into a fully validated complete system.

Agentic also does not automatically mean a higher level of autonomy; planning, feedback, and recovery are not new inventions.

## Author's Judgment (not paper content, cross-paper synthesis viewpoint)

💡 Connecting hot topics is more valuable than chasing papers in isolation
ChunkTrust focuses on when to reacquire feedback; RoboHarn-Evo focuses on how to verify results and revise experience. Put back on this map, you can see that they are respectively filling in the links of execution scheduling and experience reuse.

For engineers, the benefit of this article is: first locate which layer the bottleneck belongs to, then find methods along the corresponding direction; when evaluating combined solutions, compare success rate, time, computation, and retry budget together.