# Don't Teach the Steps: How Do Assembly Skills Emerge from Rewards?

Don't Teach the Steps: How Do Assembly Skills Emerge from Rewards?

Full article: https://haiguangboy.com/posts/alphanist

## Core Method

A shared PPO policy outputs joint targets and gripper commands based on surface point clouds of the part and mating geometry, the robot state, and the previous action.

There are no expert demonstrations, preset motion primitives, or per-task tuning. Task success reward is given only upon completing the assembly, with no intermediate rewards for approaching or aligning; action, velocity, and torque costs are still retained at each step.

Different geometries share the full assembly objective, giving the policy room to find effective manipulation methods for each.

## Key Results

↳ Drop the part and re-grasp it, exposing more of the insertion end.
↳ Use contact with the board surface to adjust the part's orientation.
↳ Temporarily park the USB-A, complete the rod, then return to insert it.

These skills were not individually specified by dedicated rewards. Continuous assembly also has no explicit planner dictating the order; all 10 pairs must be completed to receive a single task success reward.

## Comparison with Related Lines

- Opposing line · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: The authors judge that heterogeneous tasks can positively transfer through a shared terminal objective contradicts joint co-training on heterogeneous human-robot data is a structurally suboptimal approach
- Opposing line · [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`: Coordinated behavior of parking a part and returning to it appears in multi-pair assembly contradicts core claim: switching logic should be handled by a planner rather than learned by the policy—this is a structural rebuttal to "end-to-end scaling"
- Same line · [1,000 Tasks in 1 Day, Thanks to Inductive Bias](https://haiguangboy.com/posts/mt3-thousand-tasks) `learning_a_thousand_tasks_in_a_day_2026_08`: The authors judge that heterogeneous tasks can positively transfer through a shared terminal objective validates key counter-trend: decomposition methods ramp up quickly at 1–10 tasks but plateau around 50, while monolithic BC accelerates to catch up in the 10–50 range
- 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`: Coordination cases do not establish general planning or generalization to unseen tasks validates small-sample and best-of-N evaluations cannot be extrapolated to stable general capability
- 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
- Same line · [MT-WAM: How Future Supervision Serves Action](https://haiguangboy.com/posts/mt-wam) `mt_wam_reorienting_the_one_pass_predictive_representation_toward_action_generati_2026_09`: Scale ablations simultaneously change GPU, environment, and rollout batch validates the causal contributions of supervision objective, capacity, and branching structure have not yet been fully disentangled

## Boundaries

⚠️ No demonstrations does not mean no exploration support
The cached initial states cover scattered placement, in-hand, transport, and partial assembly; success-guided sampling prioritizes practicing medium-difficulty items. Keeping the binary objective but switching to uniform sampling drops evaluation from 97.4% to 84.3%.

All results come from simulation; 98.2% cannot be taken as an industrial reliability rate; covering 20 task types also does not equal completing an entire 20-piece assembly board in one go.

## Author's Judgment (Not Paper Content, Cross-Paper Synthesis View)

💡 How does diversity translate into capability?
What this paper supports is: shared objectives, task diversity, initial-state coverage, and experience allocation working together can make some hard-to-practice-alone cases solvable.

This is not yet a scaling law of "the more tasks, the smarter." What is more worth testing is: after expanding the task set, whether positive transfer persists, and whether these re-grasping, contact-adjustment, and coordination behaviors can transfer to real robots.