Fix RBY1 root_body_id type: pick success was always 0 for RBY1

[sjauhri]

Problem

PickTask success is unattainable for RBY1 — it reports 0 for every episode, even when the object is clearly grasped and lifted. This affects the RB-Y1 pick_benchmark evaluations (see allenai/MolmoBot#9, where the released MolmoBot-RBY1Multitask checkpoint reproduces Pick&Place/Opening/Door-Opening website numbers but gets 0/100 on Pick).

Root cause

RBY1BaseGroup (and the RBY1 gripper EE groups) pass model.body(...) — a _MjModelBodyViews object — as root_body_id, while every other robot view passes model.body(...).id (an int). MoveGroup stores the value as-is.

PickTask.get_info() classifies contacts by comparing body_rootid entries (numpy ints) against robot_view.base.root_body_id:

if other_root_body == self.env.current_robot.robot_view.base.root_body_id:
    robot_collision = True
else:
    non_robot_collision = True

An int == _MjModelBodyViews comparison silently evaluates to False (verified: np.int32(view.id) == view → False), so every gripper–object contact is classified as a non-robot collision. only_robot_collision can therefore never be true for RBY1, and Pick success is 0 at every step of every episode — both the at-end and the any-step/oracle metrics.

Franka pick benchmarks are unaffected because all Franka views already pass .id. PnP (receptacle-support-based) and Opening/Door (joint-position-based) don't use this comparison, which is why only Pick shows the discrepancy.

Notably task.py already contains a local workaround hinting at this ("Handle both int IDs and body view objects (some robot views store the view directly)") — this PR fixes it at the source instead.

Fix

Add the missing .id in rby1_view.py (base group + gripper EE groups), matching the convention used by the head group in the same file and by all other robot views. Swept the repo for other instances of the pattern; these two lines were the only offenders.

Note for reproduction on the MolmoBot pin: the bug exists at cd23bece (the commit pinned by MolmoBot's pyproject) as well, so this patch should be applied on top of the pin when evaluating the released RBY1 checkpoints.

[BlGene]

@sjauhri
Is this still a draft from your side? If not, I could merge it.

[sjauhri]

@sjauhri Is this still a draft from your side? If not, I could merge it.

Yes, still a draft to be tested on Pick benchmark for rby1