# 1,000 Tasks in a Day: It's All About Inductive Bias

1,000 tasks in a day, driven by inductive bias

Full article: https://haiguangboy.com/posts/mt3-thousand-tasks

## Core Method

↳ ① Decompose trajectories into alignment and interaction phases. Alignment only cares about the final pose of the end-effector relative to the target object—the path taken doesn't matter; interaction requires precise execution.
↳ ② Use retrieval instead of behavior cloning, and retrieve at test time. This differs from pre-training data retrieval and from using RGB to retrieve image-action pairs throughout—here, language plus geometry retrieves an entire trajectory before execution.

## Key Results

Controlled experiments with 3,450 real-robot trials: using only 3 demonstrations outperforms other methods that use 50. Decomposition itself is effective—with the same behavior cloning implementation, the decomposed version beats the non-decomposed one.

Thousand-task evaluation: 1,000 everyday tasks, over 400 objects, one demonstration per task, all collected within 24 hours, across 2,200 trials—78.25% on seen tasks, 68% on unseen ones.

## Related Approaches Comparison

- Opposite path · [latepost_xuhuazhe_202603](https://haiguangboy.com/posts/latepost_xuhuazhe_202603) `latepost_xuhuazhe_202603_2026_03`: Core claim: inductive bias is an asset in low-data regimes and a liability in high-data ones—this is a judgment with intersections, not a stance-taking contradicts route bet: behavior/action parts must be a unified model, opposing modular assembly
- Opposite path · [Switching logic shouldn't be learned by the policy](https://haiguangboy.com/posts/dr-lfd) `decompose_and_reorganize_planning_with_primitives_and_visuomotor_policies_learne_2026_08`: Key counter-trend: decomposition scales quickly at 1–10 demonstrations but plateaus around 50, while monolithic BC accelerates in the 10–50 range contradicts core claim: switching logic should go to the planner, not the policy—a structural rebuttal to end-to-end scaling
- Same path · [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
- Same path · [sunday_blog_20260717](https://haiguangboy.com/posts/sunday_blog_20260717) `sunday_blog_20260717_2026_07`: Key counter-trend: decomposition scales quickly at 1–10 demonstrations but plateaus around 50, while monolithic BC accelerates in the 10–50 range validates pretraining scale convergence generalization gap: 82→10→8→4→0pp, out-of-domain 14%→100%
- Same path · [World models misbehave because latent actions are polluted](https://haiguangboy.com/posts/cd-lam) `causally_debiased_latent_action_model_for_embodied_action_conditioned_world_mode_2026_07`: Ceiling of open-loop replay: once a trajectory starts, errors can't be detected or corrected mid-way validates two failure modes measured on DreamDojo: motion persists with zero actions; target-action transfer doesn't follow
- Same path · [Li Hongyang's whole-body intelligence surpasses GR's whole-body control](https://haiguangboy.com/posts/archon-whole-body-intelligence) `archon_blog_whole_body_intelligence_cn_20260712_2026_07`: Perception dependence: vision-only, single camera, no touch, relies on accurate segmentation validates four expansion dimensions: modality, body, task, failure—parameter count not among them

## Boundaries

Single-arm, single interaction, assuming the grasped object has the same pose in the gripper during demonstration and testing, and explicitly no evaluation with distractors blocking the demonstrated trajectory—open-loop replay ignores obstacles and may collide. Vision-only without touch, so deformable object stiffness and elasticity can't be inferred. Retrieval also can't interpolate between two demonstrated trajectories; it only picks one. In the appendix, the authors admit: when letting a large model chain skills, open-ended long-horizon instructions hallucinate incoherent sequences—like pouring a cocktail without putting down the bottle first.

## Author's Judgment (Not in the paper, synthesized across papers)

The most valuable part isn't the 1,000—it's the failure directions the authors proactively provide.

This judgment has intersections: the same inductive bias is an asset in low-data regimes and a liability in high-data ones. Decomposition exploits built-in task structure, so it's strong with small data; precisely because of structural constraints, scaling capacity is limited, and retrieval selects only a single demonstration, blocking cross-demonstration knowledge sharing.

This places it against both camps. Those advocating unified models and opposing modular assembly are rebutted by its low-data findings; those advocating handing switching logic to the planner are rebutted by its high-data trends. Each side is half right, depending on where you sit on the curve.

Comparing with another recent work makes this clearer: that one used 200 generated data to hit 84% success on real robots, reporting only gains throughout. This one reports gains, intersections, failure attributions item by item, and the expectation of eventually being overtaken. I discount the former's credibility; the latter, I don't.