cs.ROMar 18, 2026
SaveOffload or Overload: A Platform Measurement Study of Mobile Robotic Manipulation Workloads
Organizations: UC Berkeley · Microsoft Corp.
Abstract
Mobile robotic manipulation--the ability of robots to navigate spaces and interact with objects--is a core capability of physical AI. Foundation models have led to breakthroughs in their performance, but at a significant computational cost. We present the first measurement study of mobile robotic manipulation workloads across onboard, edge, and cloud GPU platforms. We find that the full workload stack is infeasible to run on smaller onboard GPUs, while larger onboard GPUs drain robot batteries several hours faster. Offloading alleviates these constraints but introduces its own challenges, as additional network latency degrades task accuracy, and the bandwidth requirement makes naive cloud offloading impractical. Finally, we quantify opportunities and pitfalls of sharing compute across robot fleets. We believe our measurement study will be crucial to designing inference systems for mobile robots.
Figures & tables
| Task type | Focus | Examples |
|---|---|---|
| Perception | Sense & understand the environment | SLAM, object detection, semantic mapping |
| Planning | Interpret high-level goals & plan actions | Command interpretation, task sequencing |
| Navigation | Move safely & efficiently to a target location | Path following, obstacle avoidance |
| Manipulation | Interact physically with objects | Grasping, lifting, opening, placing |
Table 1. Examples of typical tasks required to enable a mobile robotic manipulation application.
Figure 1. Robotic platforms used for evaluating workloads.
Figure 2. Experimental setup of robots, edge compute with Wi-Fi and 5G connectivity, and cloud compute.
| Class | Workload | Task description | Operations involved | Embodiment | BW | Lat | Cmp |
| Semantic understanding & long-term planning | VLMaps ( Huang et al., 2023 ) | The robot updates a semantic map with new information as it moves from point A to point B of the lab. | Perception & navigation | TurtleBot | H | M | H |
| GraphEQA ( Saxena et al., 2025 ) | The robot is asked to check if the door of the lab has been left open. | Perception, planning, & navigation | Stretch 3 | L | M | H | |
| Multi-step mobile manipulation | ( Intelligence, 2025 ) | The robot is tasked with handing over a tape from one arm to the other and to place it in a box within a time window of 70s. | Perception, planning, & manipulation | SO-101 | H | L | H |
| DreamZero ( Ye et al., 2026 ) | The robot is tasked with grabbing a frying pan from the hob. | Perception, planning, & manipulation | Simulation (video rec.) | H | L | H | |
| Collision-free navigation | RTAB-Map ( Labbé and Michaud, 2019 ) & nvblox ( Millane et al., 2024 ) | The robot is tasked to navigate to some location, which requires it to turn around a corner and avoid a previously unmapped obstacle. | Perception & navigation | TurtleBot | H | L | M |
Table 2. Robotic workloads considered for the experimental evaluation characterized by bandwidth ( BW ), latency ( Lat ), and compute ( Cmp ) requirements as low (L), medium (M), or high (H).
Figure 3. Trajectory of TurtleBot for navigation task and unmapped hidden obstacle.
Figure 4. GPU and CPU memory consumption of workloads.
Figure 5. VLMaps execution time across compute platforms.
Figure 6. GraphEQA execution time across compute platforms.
Figure 7. Nvblox success rates across compute platforms.
Figure 8. execution times and success rates across compute platforms.
Figure 9. Power consumption of onboard compute when running the workloads of Table 2 .
Figure 10. Stretch 3 battery lifetime increase from offloading, by replacing high-end onboard compute with a Raspberry Pi 5.
Figure 11. Inference time for and DreamZero across different compute platforms.
Figure 12. Example 5G cell configurations with tradeoffs in terms of latency and throughput, as measured in our testbed.
Figure 13. accuracy with different latency and bandwidth.
| Config. | 15 FPS | 10 FPS | 5 FPS | 1 FPS |
|---|---|---|---|---|
| 640x480 | 20.3 / – | 13.7 / 91.5% | 7.2 / 89.8% | 1.8 / 71.2% |
| 320x240 | 5.4 / 84.7% | 3.8 / 84.7% | 2.1 / 79.7% | 0.7 / 69.5% |
| 160x120 | 1.7 / 81.4% | 1.2 / 79.7% | 0.8 / 78.0% | 0.5 / 72.9% |
Table 3. Average bandwidth (Mbps) / recall with VLMaps across image resolutions and frame rates.
| C1. Batch Runtime (s) | C2. Speedup vs. Seq. | C3. Latency vs. Thor B=1 | C4. GPU Mem (GB) | C5. Mem. Savings | |||||||||||
| Model | B=1 | B=2 | B=4 | B=1 | B=2 | B=4 | B=1 | B=2 | B=4 | B=1 | B=2 | B=4 | B=1 | B=2 | B=4 |
| ( Intelligence, 2025 ) | 0.0633 | 0.0956 | 0.158 | 1.00 | 1.32 | 1.60 | +56.1% | +33.7% | -9.5% | 7.90 | 8.25 | 8.70 | 0% | 47.8% | 45.6% |
| SAM ( Kirillov et al., 2023 ) | 0.366 | 0.725 | 1.448 | 1.00 | 1.01 | 1.01 | +12.9% | -45.5% | -79.3% | 2.52 | 2.53 | 2.56 | 0% | 49.3% | 71.5% |
| CLIP ( Radford et al., 2021 ) | 0.0758 | 0.1398 | 0.263 | 1.00 | 1.08 | 1.16 | -10.6% | -8.7% | -31.7% | 3.90 | 3.96 | 4.08 | 0% | 49.7% | 47.9% |
| Depth | |||||||||||||||
| Anything ( Yang et al., 2024 ) | 0.0722 | 0.133 | 0.273 | 1.00 | 1.08 | 1.06 | +4.8% | -14.5% | -27.2% | 1.28 | 1.29 | 1.29 | 0% | 49.7% | 74.8% |
Table 4. Batching of robotic models on A100 GPU. (C1.) Batch Runtime is the time to process all requests as a batch. (C2.) Speedup relative to running requests one after another. (C3.) Latency vs. Thor is the latency relative to a single Thor request ( = faster, = slower). (C4.) Memory utilizations. (C5.) Mem. Savings vs. Inst. is the memory reduction compared to loading separate instances.
Figure 14. SM utilization for studied workloads on DGX Spark.
Figure 15. SM utilization for studied workloads on A100 VM.
Figure 16. Impact of and VLMaps sharing the compute.
Figure 17. Ping RTT of idle client with varying additional connected clients, each generating 50Mbps of UL TCP traffic.
More in this topic
Robotics
Robotics studies the design and control of machines that perceive, move, and act in physical environments.
See all 751 papers