REBOOT describes a bimanual assembly attempt through five shared phases, then attaches the mechanical conditions that make each phase difficult. Reaching a connector’s port, resolving its orientation and seating it are different achievements. Retaining installation/removal direction, rotational symmetry and engagement clearance makes a phase score explain where progress stops.
The REBOOT project reports 2,160 demonstrations across 18 tasks: nine objects, each with install and remove variants. Half are expert demonstrations and half are recovery episodes. This article interprets the project’s task definitions and examples; no dataset playback or assembly-policy evaluation was performed.
The five phases distinguish approach from engagement
The same phase sequence organizes all tasks, allowing failures at comparable points to be grouped without treating every incomplete rollout as the same event.
| Phase | Intended progress | Failure example from the project |
|---|---|---|
| Align(pick) | Bring the arms to the grasp configuration | Overshooting the target or stalling in free space |
| Engage(pick) | Close the grippers and secure the object | Closing at a misaligned pose and missing the object |
| Transport | Move the grasped object to the target | Drifting off-axis or dropping the object |
| Align(place) | Orient the object at the placement target | Misaligning the connector axis with the port |
| Engage(place) | Complete release and seating | Jamming partway through insertion |
The last two phases are especially useful for reading an insertion attempt. An object can arrive near the correct location while its axis or yaw remains wrong. It can then reach a plausible alignment but jam during engagement. The phase boundary separates achieving the placement alignment from completing engagement as the parts mate.
REBOOT pairs induced failures with expert recovery trajectories that return to a valid continuation state. In the project’s USB-A illustration, an offset leads the housing to catch on the port rim; the recovery retracts, realigns and approaches again. The useful training signal includes that route back into the task, alongside the original failed engagement.
Symmetry and clearance describe different demands
The task metadata labels rotational symmetry as continuous, C2 or C1. Within the project’s classification, these correspond to any valid rotation, two valid orientations or one valid orientation. Engagement clearance supplies a separate measure of the mating geometry.
USB-C is a useful example of why the two labels should remain separate. REBOOT assigns it C2 symmetry but a tight 0.09 mm engagement clearance. Two valid orientations reduce one orientation ambiguity, while the tight mating geometry still demands precise placement. The USB-A entry has C1 symmetry and a 0.62 mm clearance, so the labels describe a different combination rather than a simple ordering from easy to hard.
Those numbers are the project’s task-suite metadata. They describe its selected interfaces and fixtures; a connector family name alone does not specify the clearance of another assembly. Likewise, the suite includes a threaded M12 fastener as well as peg-in-hole mechanisms, so the engagement phase can involve different mechanics even when the phase name is shared.
The phase label says when progress failed. Symmetry, clearance and mechanism help explain which mechanical demand the policy was facing at that point.
Removal changes the starting constraint
Install and remove are matched task variants, not interchangeable trials. During installation, the policy must establish a valid relationship between initially separate mating parts. During removal, it starts with parts already engaged and works to separate them.
The project reports mean rollout success of 16% for installation and 56% for removal in its evaluated suite. It also reports that Align(place) and Engage(place) account for 54% of installation-task failures, while Transport accounts for 8%. These aggregates point toward alignment and engagement as important bottlenecks in that evaluation; they describe the authors’ task mix and tested policies.
For a concrete comparison, retain the object and operation beside the phase. “USB-C install, Align(place)” identifies a much more specific event than “placement failed.” It keeps orientation resolution and insertion clearance in view instead of blending them with extraction performance.
A phase result needs its entry population
The project’s progress analysis uses the distribution of the last completed phase. When analyzing such a dataset, an overall success fraction and a conditional success fraction can tell different stories.
Suppose, in a hypothetical example, eight of ten attempts reach Align(place) and three eventually complete the task. Overall completion is 3/10 = 30%; completion among attempts that reached Align(place) is 3/8 = 37.5%. The second number concentrates on later execution and omits the two attempts lost earlier. These are explanatory calculations, not additional REBOOT results.
A recovery can also revisit alignment after a failed engagement. Keeping that earlier event in the same task history shows the extra path required to finish, rather than presenting the retry as an unrelated clean attempt. The robot evaluation guide supplies the broader distinction between trial outcomes and intermediate progress.
REBOOT’s phase structure is most informative when the mechanical labels travel with it: which interface, which direction, which orientation alternatives, and which clearance. Together they turn an incomplete rollout into a located assembly problem.