AI2's MolmoMotion hits NeurIPS Highlight: 3D point-trajectory forecasting lifts robot pick-and-place to 76.3%

k7agar · x · 2026-10-03

MolmoMotion, a 3D point-trajectory forecasting model from AI2, UW and UNC, was selected as a NeurIPS 2026 Highlight paper with everything open-sourced.

• MolmoMotion-1M: the largest dataset of action-described, object-grounded 3D point trajectories, built from 1.16M videos
• MolmoMotion-4B: a 4B language-conditioned model that predicts each queried point's future 3D trajectory in a metric world frame from a short RGB history plus an instruction, handling rigid, articulated and deformable motion; SOTA on PointMotionBench
• PointMotionBench: a human-verified benchmark spanning 111 object categories and 61 motion types

Downstream: fine-tuned on DROID real-robot videos, initializing a manipulation policy with it lifts pick-and-place success from a Molmo 2 baseline of 56.0% to 76.3%; its trajectories also serve as explicit motion-control signals for image-to-video generation, beating the base model on all five motion metrics.

Original post →

More from Embodied

• Apple M5 Ultra 256GB Context Benchmark Shows Strong Prefill Speed on Quantized Qwen — antirez · 2026-10-03
• Dev Builds Modular MOSS Rover That Runs on Pi 5, Jetson and Any Depth Camera — OpenAIDevs · 2026-10-03
• BIND Action Head Ties Robot Actions to 2D Image Features for Data-Efficient Policies — yuewang314 · 2026-10-03
• Google's September AI recap: Gemini 4 Argon, 1M-token output, Googlebook laptops — GeminiApp · 2026-10-03
• Neuralink implants pass 50,000 hours of use, fueling a brain-computer interface foundation model — DimaZeniuk · 2026-10-03
• Microrobot navigation policies trained in under 10 minutes via 8192 parallel sims — bravo_abad · 2026-10-03

Embodied channel →