ContactBook a call
Research · Project Phasor

Brain-inspired AI, built for real hardware

Project Phasor is TwoApps' neuromorphic AI research: hardware-aware models built on spiking neural networks (SNNs), which compute with sparse, brain-like spikes instead of dense maths.

Spiking neural networksHardware-aware#6 of 19 on RoboCasa365

Phasor-m7

01 / What it is

A hardware-aware, spiking-neural-network model

Phasor-m7 is a robot-policy model from Project Phasor. It builds on Xiaomi-Robotics-1 rather than being trained from scratch.

02 / Public result

#6 of 19 on the RoboCasa365 leaderboard

56.8 overall as of 18 Oct 2026, ahead of Meta ProWAM, NVIDIA GR00T and π0.5.

03 / Efficiency

~900× more efficient than frontier models (internal benchmark)

This is our internal benchmark, not an independent result. We label it that way everywhere we use it.

RoboCasa365 leaderboard · as of 18 Oct 2026

Rank
#6 of 19
Overall score
56.8
Builds on
Xiaomi-Robotics-1

Rankings change as new models are submitted — see the live leaderboard for the current standings.

See the live leaderboard

Phasor-m7 in detail

What the result is, and what it isn't

The public leaderboard result, the base model, and how to read the efficiency figure.

  • Phasor-m7 ranks #6 of 19 on the public RoboCasa365 robot-policy leaderboard with an overall score of 56.8 (as of 18 Oct 2026).
  • It ranks ahead of Meta ProWAM, NVIDIA GR00T and π0.5.
  • Rankings change. See the live leaderboard for the current standings.

Phasor-m7 builds on Xiaomi-Robotics-1. It was not trained from scratch — the leaderboard entry names Xiaomi-Robotics-1 as its base.

~900× more efficient than frontier models (internal benchmark). The figure comes from our own testing, not from the leaderboard or an independent evaluation.

  • Spiking neural networks that only compute when a signal changes
  • Hardware-aware design, so models map onto neuromorphic and analog silicon
  • Deployment targets: Intel Loihi (via Lava) and SpiNNaker2

Publications

Peer-reviewed research (separate work from Phasor-m7)

Papers on memory mechanisms in spiking neural networks. Their benchmarks and figures are not Phasor-m7 results.

  1. Journal article · 2026

    Modality-Dependent Memory Mechanisms in Cross-Modal Neuromorphic Computing

    Authors: E. Blessing, J. Rehman, S. Roy, M. Kim, C. Tseng, I. Nkrumah

    IEEE Computer, Special Issue on Neuromorphic Computing, 2026

    The first cross-modal ablation of memory mechanisms in spiking neural networks, across visual (N-MNIST, DVS-Gesture, Caltech101-DVS) and audio (SHD, SSC) benchmarks. Hopfield memory reaches 97.68% on visual tasks but 76.15% on audio; the full hybrid has the best cross-modal average (89.99%); an optimized HGRN reaches 719× energy efficiency versus dense ANNs at 99.3% sparsity.

  2. Conference paper · 2026

    Memory-Augmented Spiking Networks: Synergistic Integration of Complementary Mechanisms for Neuromorphic Vision

    Authors: E. Blessing, C. Tseng, I. Nkrumah, J. Rehman

    NICE 2026 (Neuro-Inspired Computational Elements Conference), proceedings paper and poster, 2026

    Combines complementary memory mechanisms in spiking networks for neuromorphic vision. The full hybrid reaches 97.49% accuracy at 1.85 μJ per inference with 97.0% sparsity (silhouette 0.715); HGRN adds +1.01% accuracy and 170.6× energy efficiency.

Author lists are shown exactly as published. TwoApps co-founder Junaid Rehman is a co-author on both papers.

Next Step

Talk to the team behind Phasor

Want to know how hardware-aware AI could fit your product, or what we ship today? Book a call.

Book a callPick a time