Venue: ICLR 2026 · Authors: Hao Wang, Chengkai Hou, Xianglong Li, Yankai Fu, Chenxuan Li, Ning Chen, Gaole Dai, Jiaming Liu, Tiejun Huang, Shanghang Zhang · arXiv 2506.06690 · Category: RL for manipulation (high-speed dynamic control) · Trend tag: neuromorphic vision + imitation learning

flowchart LR
  B[Incoming ball<br/>high-speed dynamics] --> S1[System 1: fast<br/>ball detection + preliminary<br/>trajectory prediction, ms-level]
  S1 --> S2[System 2: slow<br/>SONIC spike-oriented<br/>neural calibration]
  S2 --> H[Refined hittable position]
  H --> IM[IMPACT: imitation-based<br/>motion planning + control]
  IM --> A[Robot arm striking policy]
  A --> T[Target zone: 30 cm / 20 cm]

Controlling high-speed objects in dynamic environments is a core robotics challenge, and table tennis is an ideal testbed. It poses two coupled difficulties: a high-precision vision system that can accurately predict ball trajectories under complex dynamics, and an intelligent control strategy for precise striking to target regions. Capturing rapid motion typically demands perception hardware with exceptional temporal resolution.

Inspired by Kahneman's dual-system theory, SpikePingpong integrates spike-based (neuromorphic) vision with imitation learning in a Fast-Slow architecture:

• System 1 (fast): rapid ball detection and preliminary trajectory prediction with millisecond-level responses.
• System 2 (slow): SONIC (Spike-Oriented Neural Improvement Calibrator) — spike-oriented neural calibration that refines hittable-position predictions.
• IMPACT (Imitation-based Motion Planning And Control Technology): learns optimal robotic-arm striking policies from demonstration-based learning, generating strategic hitting motions from visual input and landing information.

• 92% success rate for 30 cm accuracy zones.
• 70% success rate for the more challenging 20 cm precision targeting.

These demonstrate the Fast-Slow architecture for time-critical manipulation.

Shows that neuromorphic / spike vision paired with imitation learning can drive time-critical, high-speed manipulation where conventional frame cameras struggle with temporal resolution. The fast-slow split — cheap fast prediction plus a spike-based calibrator — is a transferable recipe for dynamic-control tasks beyond table tennis.

• arXiv 2506.06690
• OpenReview
• ICLR 2026 listing

• ICLR 2026 Survey
• RL for VLA

← Back to ICLR-2026