Caelan Garrett to present ScheduleStream, a GPU-driven multi-arm planner, at ICRA 2026 in Vienna
The researcher says ScheduleStream tackles multi-arm task-and-motion planning on GPUs; the announcement surfaced via an X post amplified by Bowen Li.
By Ryan Merket ·
Why it matters
Multi-arm manipulation is bottlenecked by planning latency. If ScheduleStream shows real GPU speedups for task-and-motion planning, it could make complex workcells and online replanning practical.
Caelan Garrett (https://x.com/CaelanGarrett) says he will present ScheduleStream, a project on GPU-accelerated multi-arm task and motion planning, at ICRA 2026 in Vienna. The announcement surfaced on X and was amplified by Bowen Li (https://x.com/Bw_Li1024).
In Garrett's words, ScheduleStream is "our work on GPU-accelerated multi-arm task and motion planning" shared ahead of the IEEE International Conference on Robotics and Automation. The post did not include a paper link, preprint, or project page, and it did not specify co-authors or affiliations. Still, the teaser hints at a timely research direction: pushing task-and-motion planning (TAMP) for coordinated robot arms onto GPUs to hit real-time or near-real-time constraints that CPUs struggle to meet.
What ScheduleStream appears to tackle
Multi-arm TAMP asks robots to decide both what to do and how to do it when more than one manipulator must share space, tools, and timing. That means juggling action sequencing, collision-free trajectories, handoffs, and resource contention, often under tight latency budgets. Researchers typically balance global task search with local motion planning and scheduling; getting that orchestration to run fast enough for physical systems is the hard part. A GPU-accelerated approach suggests ScheduleStream seeks parallelism across candidate task schedules, motion rollouts, or collision checks to shrink planning time as the number of arms and objects grows.
While there are several academic and open-source planners in the TAMP ecosystem, the X post does not make comparative claims or share benchmarks, so it is too early to place ScheduleStream on a performance leaderboard or to assess generality across robot platforms. The Vienna venue matters, though: ICRA is one of the flagship conferences for robotics systems research, and a public presentation there will put details, demos, and scrutiny in front of the field.
What we do and do not know
Based solely on the announcement:
- Topic: GPU-accelerated multi-arm task-and-motion planning.
- Venue: ICRA 2026 in Vienna.
- Messenger: Caelan Garrett, with the news surfaced via a retweet from Bowen Li.
Missing from the post are the paper title and author list, an abstract or technical summary, metrics like speedups over CPU baselines, task success rates, or the number of robot arms and scenarios tested. There is also no code or dataset link, and no indication of whether the work targets a specific robot stack, simulator, or GPU architecture.
Why founders and operators should care
If GPU-first planning makes multi-arm coordination dramatically faster, it could unlock denser workcells, tighter human-robot collaboration windows, and more fluid handoffs on factory floors and fulfillment lines. Even modest latency reductions can move TAMP from an offline planning stage to something closer to online replanning when disturbances occur. For robotics startups building manipulation systems, a production-grade GPU planner could change the bill of materials calculus and open the door to higher throughput without rewriting the entire stack.
For now, the community will be watching for the ScheduleStream paper, videos, and code to see how the approach composes scheduling with motion planning and where the parallelism pays off.