Habitat 2.0: Training home assistants to rearrange their habitat

Appendix F.2 (pp. 25–27) describes a separate Navigate skill that brings the robot to the starting position of other skills; Pick and Place instead start facing their targets with only small position and orientation variations. Section 6.2 (around p. 12) says learning navigation to chain skills remains challenging. The paper therefore does not describe shifting mobility into the manipulation subpolicies to resolve the hand-off problem. (LLM, conf=0.87)

Checked by a reasoning model — may be wrong.

Section 3 (p. 4) says static 3D scans cannot be moved or manipulated, supporting the claim that scanned scenes are less interactive. The paper also reports 900+ person-hours of artist effort for ReplicaCAD, but it does not establish that previous work mainly relies on artists or that scanning is used specifically to reduce human labor. (LLM, conf=0.82)

Appendix F.2 (p. 25) says skills must return the arm to a resting position to facilitate transitions, and its Navigate description (p. 27) says navigation brings the robot to the start of other skills. These passages support the concern that a subsequent skill needs a suitable starting state, but they do not establish that naive chaining of arbitrary skills often produces hand-off failures. (LLM, conf=0.79)

The H2.0 description and simulator comparison table support the rigid-body-physics claim, and Section 6 (p. 9) describes three HAB task families. The action-space description supports motor-controlled arm execution, but it also says grasping is abstracted: nearby objects snap into the gripper and cannot slip. Section 4.1 further says the robot base moves kinematically rather than through simulated wheel-ground contact, so the unqualified claim of realistic action execution overstates this source. (LLM, conf=0.83)

In the conclusion (p. 12), the authors say they defined the benchmark’s rearrangement tasks, supporting the point that researchers designed these tasks. Section 6 (p. 9) calls them common household assistive tasks, but the paper does not say it is unclear whether they address actual human needs or make that claim about existing benchmarks generally. (LLM, conf=0.84)

In §6.2 (p. 11), the paper reports that its monolithic RL approach performed poorly on long-horizon Home Assistant Benchmark tasks and that chaining skills without an architectural bias remained out of reach. It studies a hierarchical planner with parameterized skills, but does not establish that action-space abstraction is necessary or test BEHAVIOR-1K activities. (LLM, conf=0.86)

The simulator comparison table describes earlier Habitat as using 3D scans with no physics support, consistent with scanned objects not being interactive. But the paper does not identify HM3D scenes or specifically state that their existing objects are non-interactive in Habitat-sim. (LLM, conf=0.82)

The introduction describes Habitat 2.0 as a photorealistic 3D simulator, and Table 1 lists its rendering as “3D scans + PBR shading.” Section 5.1 (p. 7) describes a Fetch robot as the agent embodiment, but the passages found do not establish support for multiple robot embodiments. The source therefore supports the rendering part of the claim, not the multiple-embodiments part. (LLM, conf=0.83)

Section 6 (p. 9) describes a Home Assistant Benchmark of household manipulation tasks, not an ObjectNav benchmark. Searches for ObjectNav and the claimed total of 16 object types found no passage assessing the environments or combined object types of existing ObjectNav benchmarks; the paper’s 92 ReplicaCAD objects refer to a different dataset. (LLM, conf=0.73)

Section 5.1 (p. 7) describes a robot picking up objects, and the paper describes an abstracted grasp action rather than results for handle detection or grasp-pose detection. Searches for YOLO, handle detection, grasp-pose detection, and a public handle-specific dataset found no passage supporting the attributed claim. (LLM, conf=0.76)