Discord

Checklist NOAI China 2026 Round 2 (China Stage) · Task 2

Embodied Intelligence Sim2Real State Prediction Based on NVIDIA Isaac Sim

English title: 基于 NVIDIA Isaac Sim 的具身智能 Sim2Real 状态预测

Predict a real robot's 6-dimensional joint states for hidden frames from the full simulated trajectory and a visible prefix of real states.

  • Time series
  • Multivariate sequence regression

The task

Simulators such as NVIDIA Isaac Sim generate robot trajectories cheaply, but simulated and real trajectories do not align perfectly. For each task trajectory the contestant receives, for every frame, the 6-dimensional simulated joint position (simulation_positions) and the timestamp, together with the real robot's 6-dimensional joint state (observation_state) for a visible prefix at the start of the trajectory.

The goal is to predict observation_state for all subsequent hidden frames. Any approach may be used, including regression, feature engineering, retrieval of similar trajectories and nearest-neighbour methods; the statement suggests studying the relation between simulated and real trajectories and how the visible prefix informs the hidden segment. The problem was contributed by NVIDIA.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
CSV files with columns index, task_index, frame_index, timestamp, simulation_positions and observation_state (6-value arrays). Training set: 336 tasks, 176,675 rows, all states given. Validation: 72 tasks, 35,136 rows, 24,562 to predict. Test: 72 tasks, 37,632 rows, 26,310 to predict. Validation and test data are read through environment variables at evaluation time.
You submit
A notebook submission.ipynb that runs end to end and writes submission.zip with submission_val.csv and submission_test.csv, containing exactly the hidden rows with columns index, task_index, frame_index, timestamp, observation_state (a length-6 numeric array string). Only the notebook may be submitted.
Scoring
For each hidden frame, RMSE over the 6 dimensions; row score = exp(-10 * RMSE); final score = mean row score (range (0, 1]). Leaderboard A: validation set; leaderboard B: test set. Baseline B score 0.5647; Scientific Committee reference B score 0.7861.
Rules
  • CPU; training plus inference must not exceed 25 minutes.
  • No external data; no external LLM APIs; no internet access or pip install during evaluation.
  • The program must not access or recover the true labels of hidden frames.
  • Evaluated in the noai:2026v1.1 image.
Format
NOAI 2026 China Stage (Round 2), Task 2. Scheduled for 21 June 2026 (one day). Republished on Bohrium for practice (paid automatic grading).

Details

Year
2026, Beijing, China
Round
Round 2 (China Stage) · Task 2
Language
English
License
Not stated by the source