My vision is to bridge art and technology.
We're at the edge of a new realm of entertainment, built not just on LLMs but on deeper fields like computer vision and reinforcement learning.
The future of creative experience lies in systems that perceive, adapt, and respond, not just generate.
A venture building large-scale 3D data for both embodied AI and creative work. For robotics, the pipeline produces structured, physically grounded datasets, including articulated objects with movable joints exported as URDF, to train models that perceive and manipulate the physical world. For creative fields, it delivers production-ready 3D assets for games, film, and virtual worlds. More coming soon.
Visit Site →Can a frozen, general-purpose video-VAE latent be decoded directly into an explicit 3D Gaussian Splatting asset, supervised by multi-view render loss alone? On Objaverse-LVIS orbit renders, a strict ray-anchored decoder collapses into translucent fog, while a scaffolded RGBD fusion recovers ~21 dB and ~0.96 alpha-IoU on held-out objects. The latent isn't the wall: the scaffold recovers ~9 dB of lost detail, so the real bottleneck is the decoder's representation and the data, not the frozen video latent.
View Project →Behavioral-cloning policies plateau from compounding distribution shift; RL fine-tuning can correct them but is bottlenecked by sparse, binary task-completion rewards. I extract waypoints from the BC encoder's own latent space to produce an unsupervised dense reward, then run PPO on top. The residual pipeline yields +7 points over BC on LIBERO-Spatial and trains 5x faster than sparse PPO while preserving pretrained behavior.
View Project →Applied Low-Rank Adaptation to Meta's V-JEPA2-AC video world model for robotic manipulation on DROID-100. Measured energy landscapes before and after fine-tuning across 30 held-out trajectories. LoRA caused representation collapse: flat, non-discriminative landscapes, despite the technique's success on LLMs. A reminder that parameter-efficient methods designed for discrete token prediction don't transfer cleanly to continuous-valued latent spaces.
View Project →