Radar Brief Week 16, 2026 · 2026-03-04 — 2026-03-11

NVIDIA Releases 600-Hour Robotic Manipulation Dataset
AI Data Intelligence Weekly

This week scanned 86 HF orgs · 50 GitHub orgs · 71 blogs · 125 X accounts

0
Valuable Datasets
0
Related Papers
0
Blog Posts
0
Active Repos
One-line Summary

NVIDIA releases 600-hour robotic manipulation dataset, physical AI data demand surges [P0], Allen AI releases research assistant citation tracking data, Agent tool data becomes new hotspot [P0], Anthropic releases economic impact index dataset, AI application evaluation becomes new demand [P1]. This week's strongest data demand signal: robotic manipulation trajectories.

Key Findings

This week's 5 high commercial value findings

P0 NVIDIA Releases 600-Hour Robotic Manipulation Dataset, Physical AI Data Demand Surges [P0]

NVIDIA released the PhysicalAI-Robotics-Kitchen-Sim-Demos dataset on 2026-02-10, containing 600 hours of human teleoperation demonstrations covering 316 different tasks with 55k trajectories total. Concurrently released PhysicalAI-Robotics-NuRec (50 likes) and Arena-GR1-Manipulation datasets form a complete robotic training data ecosystem.

Business Implications → Physical AI is becoming the next data-intensive track. Unlike pure language models, robotic data must be collected through human demonstrations, with each task requiring precise judgment from professional operators. This creates new opportunities for Knowlyr in the "human-machine collaborative data collection" domain.
P0 Allen AI Releases Research Assistant Citation Tracking Data, Agent Tool Data Becomes New Hotspot [P0]

The allenai/asta-summary-citation-counts dataset (released 2025-10-08, 456 downloads) tracks the most cited papers on the Asta research platform, reflecting AI Agent knowledge preferences in actual usage. This is the first public "Agent usage behavior" dataset.

Business Implications → The Agent era requires new types of evaluation data — not testing whether Agents can complete tasks, but evaluating how they use tools and knowledge. Such data must be based on human judgment in real usage scenarios; traditional synthetic methods cannot generate it.
P1 Anthropic Releases Economic Impact Index Dataset, AI Application Evaluation Becomes New Demand [P1]

Anthropic/EconomicIndex (released 2025-02-06, 11,995 downloads, 473 likes) provides insights into AI integration into actual tasks in the modern economy, including labor market impact and job exposure analysis. This is the first public dataset to systematically evaluate AI's economic impact.

Business Implications → As AI deployment accelerates, enterprises and governments need to assess its socioeconomic impact. Such evaluation data heavily relies on domain expert judgment and cannot be automatically generated through algorithms, creating a new market for high-value human judgment.
P1 Google DeepMind Releases African Language Speech Dataset WaxalNLP [P1]

google/WaxalNLP (released 2026-01-19, 10,345 downloads) is a large-scale multilingual speech corpus supporting automatic speech recognition and text-to-speech tasks. The dataset uses cc-by-sa-4.0 license, demonstrating focus on low-resource languages.

Business Implications → Multilingual data collection requires deep involvement of local language experts, with speech labeling for each language requiring professional judgment from native speakers. This validates that "linguistic diversity" remains an irreplaceable domain for human judgment.
P2 Preference Learning Research Surges, 5 RLHF-Related Papers Published Simultaneously [P2]

This week saw publication of 5 RLHF papers including ActiveUltraFeedback (active learning for optimizing preference data collection), wDPO (robust preference optimization), DARC (divergence-aware alignment). Research focus has shifted from "how to align" to "how to efficiently obtain high-quality preference data".

Business Implications → Academia recognizes that preference data quality is the RLHF bottleneck and is exploring active learning, divergence handling, and other methods. These approaches all emphasize the critical role of human judgment, particularly in handling annotator disagreement and noisy data.

Demand Signals

Infer training data demands from model releases

Data Type Intensity Trend Related Signals
Robotic Manipulation Trajectories
Extremely Strong ↑ New
NVIDIA releases 600-hour demonstration data, 19 robotic datasets active
Agent Tool Usage Logs
Strong ↑ New
Allen AI releases citation tracking data, evaluating Agent behavior patterns
Preference Alignment Data
Strong ↑ New
5 RLHF papers focus on data quality and divergence handling
Multilingual Speech Data
Medium ↑ New
Google releases African language dataset, low-resource languages gain attention
AI Impact Assessment Data
Medium ↑ New
Anthropic economic index receives 11,995 downloads
Medical Professional Data
Medium ↑ New
InternLM endoscopy data shows vertical domain demand
Video Understanding/Tracking Data ↓ Dropped Present in previous issue, absent this issue
Coding Agent Trajectories ↓ Dropped Present in previous issue, absent this issue
Robotic Tactile Data ↓ Dropped Present in previous issue, absent this issue
Model Alignment Evaluation ↓ Dropped Present in previous issue, absent this issue
Professional Domain Reasoning ↓ Dropped Present in previous issue, absent this issue
CAD Design Instructions ↓ Dropped Present in previous issue, absent this issue
Audio-Visual Coordination ↓ Dropped Present in previous issue, absent this issue
Medical Dialogue Privacy ↓ Dropped Present in previous issue, absent this issue

Download Movers

Datasets with the largest download changes this week

Dataset Downloads Weekly Growth
lerobot/berkeley_cable_routing 1,784 +19.9%
lerobot/aloha_static_fork_pick_up 1,249 +12.9%
google/WaxalNLP 10,345 +2.3%
Anthropic/EconomicIndex 11,995 +1.4%
lerobot/berkeley_gnm_recon 1,194 -25.6%

Deep Dive — DataRecipe

This week's 3 high-value datasets reverse-analyzed (auto-generated by DataRecipe)

togethercomputer/CoderForge-Preview
300 samples · 7 fields · Hard
6.0/10
🟢 Recommended for replication

Data Structure

trajectory_id finish_reason image messages reward tools license

Risk Assessment

Medium Risk Labeling quality may fluctuate → Establish strict QA processes, set quality thresholds
Low Risk Data may become outdated over time → Establish continuous update mechanism
allenai/Dolci-Think-SFT-32B
300 samples · 3 fields · Hard
6.0/10
🟢 Recommended for replication

Data Structure

messages id source

Risk Assessment

Medium Risk Labeling quality may fluctuate → Establish strict QA processes, set quality thresholds
Low Risk Data may become outdated over time → Establish continuous update mechanism
google/MapTrace
300 samples · 3 fields · Medium
6.5/10
🟢 Recommended for replication

Data Structure

image input label

Risk Assessment

Medium Risk Labeling quality may fluctuate → Establish strict QA processes, set quality thresholds
Low Risk Data may become outdated over time → Establish continuous update mechanism

This week analyzed 3 datasets · Human ratio 99.6%

Want to discuss this issue?

Kai
Kai Founder & CEO
苏文
苏文 AI Documentation & Release Engineer
陆明哲
陆明哲 AI Product Manager

Auto-generated by AI Dataset Radar · Updated weekly

AI Dataset Radar →