# CBST: What Should Be Preserved in Cross-Embodiment Transfer

CBST: What Should Be Preserved in Cross-Embodiment Transfer

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

## Core Method

The authors propose CBST (Causal Bundle-guided Subtask Transfer), framing the problem as "which operational information should be preserved at the current moment."
↳ ① Learn representations first: pretrain on the 174 action classes of SSv2 for temporal classification, then ground the representations on robot demonstrations onto conditions, manipulation attempts, object changes, and local effects.
↳ ② Select conditions: within a predefined manipulation model, choose a set of variables that maintains the current mechanism's output and still responds to changes in the retained input combinations, updating with the execution context.
↳ ③ Connect to WAM: convert the selected variables and values into text conditions and feed them to a frozen world action model. Actions are generated by the robot using its own observations; when selection is infeasible, the corresponding structural fields are omitted while task and state conditions are kept.

## Key Results

· RoboTwin Easy: under the same robot adaptation budget, ordered human pretraining yields 94.00%, versus 92.80% without it.
· With the same CBST checkpoint, omitting the selected dependency-related fields gives 92.82%, versus 94.00% with full conditions.
· AgileX real-robot cup placement: 14/20→17/20. Only the final placement after releasing the gripper counts as success; the dual-arm handover is merely a process step.

## Comparison with Related Lines of Work

- Opposite 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
- Opposite line · [MoCaP: Human Video-Driven Zero-Shot Mobile Manipulation](https://haiguangboy.com/posts/mocap) `mocap_mobile_code_as_policies_for_zero_shot_mobile_manipulation_from_passive_hum_2026_10`: what is frozen is the WAM, not the entire method requiring no training contradicts all learning components frozen, no task-specific training
- Opposite line · [Don't Teach Steps: How Do Assembly Skills Emerge from Rewards?](https://haiguangboy.com/posts/alphanist) `alphanist_emergent_assembly_skills_from_binary_task_rewards_2026_10`: the authors stress that the same goal does not mean the implementation conditions are preserved contradicts the authors' judgment that heterogeneous tasks can transfer positively through a shared terminal goal
- Same line · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: human video first trains action classification representations validates joint co-training on heterogeneous human-robot data is a structurally suboptimal solution
- Same line · [DexTouch-WM: Human Tactile Pretraining](https://haiguangboy.com/posts/dextouch-wm) `dextouch_wm_learning_action_conditioned_tactile_world_models_from_human_touch_fo_2026_09`: the authors argue for transferring effect-related conditions and dependencies validates the authors' claim: after sharing the contact interface and action space, human touch can become a scaling axis for robot prediction
- Same line · [KODA: Converting Human Demonstrations into Humanoid Whole-Body Data](https://haiguangboy.com/posts/koda) `koda_kinematics_observation_data_alignment_for_humanoid_learning_from_human_egoc_2026_10`: a shared semantic template does not guarantee that human beams are effective on robots validates there is still tracking residual between reusing images and the implementation state

## Boundaries

⚠️ Boundaries and Costs
"Causal validity" is limited to the predefined model and the estimated context; if the template or state estimation is wrong, condition selection may also fail. Robot demonstrations, semantic prediction, and selector training are still required. The real-robot numbers above are only 20 executions each for the cup task; the public baselines in the main simulation table were not rerun in a paired manner, so small differences cannot be directly treated as significant gains.

## Authors' Judgment (Not Part of the Paper; Cross-Paper Synthesis)

Viewed alongside MoCaP: it transfers search and approach priors, CBST filters local manipulation conditions, and the concrete actions are all left to be generated on the robot side. The key to cross-embodiment reuse lies in "which experiences can be borrowed and which execution conditions need to be re-grounded." Looking back in six months: with the same robot data and models, can structural selection consistently outperform goal/state conditions in cross-task re-evaluation?