Testing the assumption "more diverse is better" across 3 axes. Task diversity matters more than many demonstrations /task. Multiembodiment pretraining is optional & human demonstrator diversity can hurt, w/velocity multimodality the culprit
ieeexplore.ieee.org/document/114889…
![We investigate critical aspects of data diversity for robotic manipulation systematically, i.e., task, embodiment, and expert diversity. Through comprehensive evaluation in simulation and the real world, we reveal key insights that challenge conventional assumptions on data scaling. (a) Task diversity benefits policy learning with predictable power– law scaling. (b) Multiembodiment pretraining data are nonessential for cross-embodiment transfer capabilities—models pre-trained on high-quality single-embodiment data can efficiently adapt to different embodiments and show desirable scaling properties during fine-tuning, which can serve as a competitive alternative to large-scale multiembodiment pretraining. (c) Expert diversity confuses robot learning, toward which we devise a distribution debiasing method based on GO-1 [1]; the yielding GO-1-Pro attains superior data efficiency during both pretraining and finetuning, where it achieves substantial performance gains of 15%, equivalent to using 2.5 times the pretraining data. Project page: https://github.com/OpenDriveLab/AgiBot-World.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:ttw5tcvxt52gbygtxrgvschk/bafkreidyofwfqhs3hvebtzg2lu7atpybeqcqdprs4hmre5lxwr7bs5ybxu)
![Deployment of the proposed detection and tracking algorithms as part of a cooperative navigation multirobot system in various environments [52]. (a) Warehouse environment. (b) Forest environment.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:ttw5tcvxt52gbygtxrgvschk/bafkreiayfxmn5lnt5m2atqx7hiv63xzbfs3sgwbnxdphq5ao5jkd46zneq)
![Illustration of the general framework, in which the dashed box highlights our contribution. Given the current partial object model and a set of candidate views, an NBV algorithm (RSV) [1] selects the next view to visit. Our method then uses this target NBV and the current partial model to select an informative focus point and employ a whole-body control strategy to reach the target while keeping the focus point within the camera FoV with a visibility constraint [11], avoiding obstacles with VFIs [12], and mitigating local minima at obstacle boundaries with a circulation constraint [13].](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:ttw5tcvxt52gbygtxrgvschk/bafkreieramgx4vdzffmrac6vaitkafet2wzvgh5tm5bjdwtu5vvf25l6om)
![Representation of bilevel optimization procedure for the robotic prosthesis. Left panel: The robotic knee torque, τ , is generated from the impedance control parameter values, which are obtained through real-time tuning by the RL controller. Lower right panel: This panel illustrates features of knee and thigh kinematics throughout a single gait cycle. For knee kinematics, superscript numbers 1–4 denote the respective phases within the gait cycle (i.e., STF, STE, SWF, and SWE), each corresponding to a knee feature. For thigh kinematics, the minimum value of the thigh angle acts as the feature. Top right panel: At the end of each bilevel optimization iteration, the cost function derived from IRL, each in a quadratic form, is utilized in the design of the RL controller. The implementation of the two interleaving procedures involving the inverse RL and forward RL is summarized in Algorithm 1. The RL controller’s inputs include kinematic features from the corresponding phase [defined in (5) and (6)], and its outputs involve adjustments to the impedance settings [defined in (3)].](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:ttw5tcvxt52gbygtxrgvschk/bafkreicouj2gktsqqrlrflv7z4kwavlcomzy2uso32cntiuhx6fj3s3hdu)
