Mem4Act: Learning Effective Latent Memory for Long-Horizon Robotic Manipulation

Name: Mem4Act: Learning Effective Latent Memory for Long-Horizon Robotic Manipulation
Uploaded: 2026-10-19T18:42:21.986069+00:00
Duration: 546 s
Description: Mem4Act introduces a framework to learn fixed-size latent memory for long-horizon robotic control. It tackles coverage, content, and utilization challenges using multi-granularity temporal aggregation for history organization, semantic/perceptual supervision via QA and DINO features, and a memory-first two-stage training strategy. This results in high success rates on complex manipulation tasks.

Mem4Act introduces a framework to learn fixed-size latent memory for long-horizon robotic control. It tackles coverage, content, and utilization challenges using multi-granularity temporal aggregation for history organization, semantic/perceptual supervision via QA and DINO features, and a memory-first two-stage training strategy. This results in high success rates on complex manipulation tasks.

Open the full episode