# RoboHarn-Evo: Knowledge-Layer RSI

RoboHarn-Evo: Knowledge-Layer RSI

Full article: https://haiguangboy.com/posts/roboharn-evo

## Core Method

Knowledge guides action, physical feedback revises knowledge, which is then reused for subsequent tasks. But this loop cannot only store "success or failure."

A robot grabs a cup but places it off-target; the grasping method may be correct, yet the placement task still fails. The overall success or failure cannot tell the system which step to keep and which step to correct.

↳ The task layer records what to do next and when it counts as complete; the action layer records the approach direction relative to the object, the interaction position, and the expected effect. The two layers are verified separately, then linked for reuse.

↳ During execution, a subtask is selected first, then action strategies constrained by the goal are retrieved; the specific pose is recomputed based on the current scene, and what is transferred is the strategy and its applicable conditions.

## Key Results

Only when the knowledge is applicable and the corresponding strategy is actually executed is the result recorded as support, opposition, or unverified. Subsequent observations from the same attempt complete the original judgment without double counting.

Counterevidence causes the old version to exit retrieval; a revised strategy still requires matching execution evidence before it can be reused. Writing a more reasonable explanation does not mean the error has already been fixed.

## Comparison with Related Lines

- Opposing line · [Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design](https://haiguangboy.com/posts/transformer-transformer-codesign) `transformer-transformer_blog_transformer_transformer_a_unified_20260807_2026_08`: Fixed model and actuator, updating task-action two-layer knowledge contradicts the line's claim: task failure should not be attributed to the policy by default; body-controller-policy should enter the same optimization problem
- Opposing line · [latepost_xuhuazhe_202603](https://haiguangboy.com/posts/latepost_xuhuazhe_202603) `latepost_xuhuazhe_202603_2026_03`: Fixed model and actuator, updating task-action two-layer knowledge contradicts the line's bet: the behavior/action part must be a unified model, opposing modular assembly
- Opposing line · [Robot GPT-3 Moment: Plenty of Story, Not Enough Evidence](https://haiguangboy.com/posts/gen-1-5-one-shot) `generalist_blog_gen_1_5_20260818_2026_08`: Self-improvement is limited to knowledge content and retrieval eligibility, not self-rewriting of the model or controller contradicts the core narrative (company self-report, not third-party verified): these capabilities emerge directly from scaled pretraining and require no specialized mechanism design
- Same line · [Agile-WAM: Lightweight Visuotactile Joint Prediction](https://haiguangboy.com/posts/agile-wam) `agile_wam_an_agile_tactile_world_action_model_for_contact_rich_robot_control_2026_09`: Real-robot only three tasks with 8 scenes each, π0.5 baseline not fine-tuned on real robot validates that small-sample and best-of selection evaluation cannot be extrapolated to stable general capability
- Same line · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: Combining task and action knowledge outperforms using each alone validates MM-DiT action expert: joint global attention fusion outperforms naive DiT
- Same line · [LAWM: Why Action Labels Become a Burden](https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling) `latent_action_pretraining_through_world_modeling_2026_07`: LAWM: Why Action Labels Become a Burden

## Boundaries

⚠️ Not yet the endpoint of long-term reliability
The authors call it RSI; the direct evidence is behavioral improvement brought by knowledge updates, and it has not yet separately proven that the improvement capability itself continues to strengthen. The model and execution primitives remain fixed; error handling includes repair and deactivation, knowledge correctness is audited by an external GPT-6, and the scale of real-robot validation is limited.

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

💡 Comparison with other improvement lines
PAVE computes advantage per action chunk, avoiding judging remedial actions in failed tasks as bad along with everything else, but each round still requires retraining the policy. This work instead writes the gains from experience into external knowledge.

SmoothRL requires optimization to align with the actions actually executed; this work requires evidence to correspond to the strategies actually executed. The common question is: after failure, where exactly should the system change so that the next execution is more reliable.