$ cat wiki/papers/2025/2511.19430-grant-parallel-task.md
Cook and Clean Together: Teaching Embodied Agents for Parallel Task Execution (GRANT)
TL;DR
Embodied agents execute instructions serially even when the physical world permits overlap — cleaning the sink while the microwave runs. GRANT adds a scheduling token to an embodied multimodal LLM so the model emits a parallel schedule alongside its grounded actions, cutting total completion time by 30.53%.
Authors & Org
Liang, Zhang, Xu, Ju, Luo, Bai. Code released under the H-EmbodVis organization
(source).
Accepted as an AAAI 2026 Oral. Posted to arXiv November 2025 — this is not a new preprint. It reached #1 on HuggingFace Daily Papers on 2026-07-30 (113 upvotes), which is why it is recorded here: a curation event, not a publication event.
Method
The paper defines ORS3D — Operations Research knowledge-based 3D Grounded Task Scheduling — a task requiring three capabilities at once: language understanding, 3D grounding, and efficiency optimization. The objective is minimizing total completion time by exploiting parallelizable subtasks.
GRANT is an embodied multimodal LLM with a scheduling token mechanism that generates the schedule and the grounded actions together, rather than planning and grounding in separate stages.
Dataset: ORS3D-60K — 60K composite tasks across 4K real-world scenes.
Results
| Measure | vs. baseline |
|---|---|
| Time Efficiency | +30.53% |
| 3D grounding accuracy | +1.38% |
| The asymmetry is the finding: grounding barely moves, scheduling moves a lot. The gain comes | |
| from reordering work the agent could already do, not from perceiving the scene better. |
Significance
Most embodied-agent benchmarks score whether a task completes. This one scores how long it takes, which turns wall-clock into a trainable objective and makes idle time a modelled cost. That matters more as the acting layer improves: Gemini Robotics 2 reports whole-body manipulation success from 45.7% to 76.3%, and once individual actions are reliable, the remaining headroom in a multi-step household task is scheduling.
It also lands on the same seam as Gemini Robotics ER 2, where a reasoning layer orchestrates over an acting layer — GRANT is that orchestration problem stated as a benchmark with a time objective.
Open Questions
- Whether the +30.53% Time Efficiency figure is simulation-only. The abstract-level material read here does not say (source).
- Model size and base model are not stated in what was read.
- ORS3D-60K is built over "real-world scenes", but whether execution is evaluated on hardware is not established.
- Whether the scheduling token generalizes past the household-task domain the dataset covers.
Cite
Liang, Zhang, Xu, Ju, Luo, Bai.
Cook and Clean Together: Teaching Embodied Agents for Parallel Task Execution.
arXiv:2511.19430. AAAI 2026 (Oral).