paper_rob__bpp

What problem does it solve?

Conditioned on semantically salient keyframes detected by a visual-language model (VLM), BPP reduces history distribution shift in long-horizon robot manipulation by remembering and leveraging only task-relevant events.

Core Features & Use Cases

• Independent 1 Hz keyframe detection with latency masking (Δ = 3 s) to simulate deployment latency.
• Multiview perception with separate ResNet34 encoders per camera, a transformer-based policy backbone, and a 50-step DDPM action head.
• Flexible prompts and integration guidance enabling plugging keyframe conditioning into existing diffusion transformer policies for real- and simulated tasks (ALOHA 2, MuJoCo).

Quick Start

Start by configuring a 1 Hz VLM keyframe detector with a 3-second latency mask and plug its keyframes into a diffusion-transformer policy to generate 50-step action chunks at 50 Hz.