Skip to content

Low rollout efficiency with frequent planning retries on RTX5090 and isaacsim-5.1.0 #6

Description

@TorwnexialX

Hi, and thank you for releasing this great work! We have been trying out InternDataEngine for a potential large-scale rollout and would appreciate your guidance on whether our observed throughput and failure patterns are expected under the published task configurations.
Across more than 100k plan calls, the overall plan success rate was 43.4%, while persisted trajectory time represented only 20.5% of the provisioned GPU time. We also observed strongly task-dependent IK/trajectory failures, along with recurring camera mounting/API compatibility warnings.

We are using the pipe mode provided in the codebase, with 1 plan worker and 1 render worker per GPU. Our environment details and observations are described below.

Environment and hardware

Component Value
GPUs 4 × 8 RTX 5090
OS Ubuntu 24.04.2 LTS; kernel 5.10.134-013.5.kangaroo.al8.x86_64
NVIDIA driver 570.153.02; CUDA 12.8
Isaac Sim 5.1.0
CuRobo 0.7.4.post1.dev1+dirty
Python / PyTorch Python 3.11.15; PyTorch 2.7.1+cu128 (CUDA 12.8)

Note: We used Isaac Sim 5.1.0 because we encountered rendering issues with Isaac Sim 4.1.0 on RTX 5090 GPUs.

Core issue

The main concern is low compute-to-product efficiency. Across the 36-hour observation window, persisted trajectory time amounted to only 20.5% of provisioned GPU time, while 42.2% of planning time produced no sequence. Because persisted episodes are counted before camera and other downstream quality filtering, the true usable product/input ratio may be lower.

Metric Observed value
Plan success rate 43.4% across more than 100k plan calls
Persisted trajectory / provisioned GPU time 20.5%
Trajectory produced per GPU-hour 12.3 min
Persisted trajectory / planning time 30.0%
Planning time producing no sequence 42.2%

Approximately 70.7% of failed plan calls were attributed to IK/reachability failures (MotionGenStatus.IK_FAIL or KPAM_IK_NO_SOLUTION). We are not yet sure about the root cause, but our current hypothesis is that some combinations of randomization ranges and keypoint constraints may occasionally lead to unreachable targets with our current Isaac Sim setup. We would appreciate confirmation of the expected throughput and the recommended task- and embodiment-specific settings.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions