Robotics Breakthroughs: RoboTTT, ASPIRE, ENPIRE

NVIDIA GEAR lab, led by Dr. Jim Fan, announced three robotics breakthroughs in mid-2026. RoboTTT (Test-Time Training) scales robot context to 8,000 timesteps (~5 minutes) with constant inference cost by embedding a small neural network that compresses history via gradient steps. It enables one-shot imitation from human video, on-the-fly error recovery, and a context scaling curve with 62% improvement from 1K to 8K context. ASPIRE gives robots a self-evolving skill library: coding agents observe multimodal traces, launch evolutionary search over control programs, and distill know-how into an ever-expanding repository. It achieves up to 10x reduction in transfer learning tokens for sim-to-real and cross-embodiment transfer, with 150+ tasks and 90+ skills demonstrated. ENPIRE deploys 8 Codex agents with a robot fleet and GPU allocation to autonomously solve physical tasks. The agents look for visual clues, reset scenes, practice skills, read papers, debate, and iterate directly on hardware. A 'physical scaling' effect was observed: 8 robots in parallel improve faster than fewer. All three systems will be open-sourced.
Key insights
RoboTTT enables robots to learn from long context and improve continuously after deployment
By performing test-time training on every sensor reading, RoboTTT compresses history into a fixed-size hidden state, allowing 8,000 timesteps of context with no inference cost increase. This supports one-shot imitation from human video, mid-episode error recovery, and a context scaling curve that shows no saturation, suggesting even 1M context is feasible.
ASPIRE replaces weight updates with an evolving skill library for continual robot learning
Instead of gradient descent, ASPIRE uses coding agents to search over control programs, distilling successful strategies into a library. This library transfers across simulation and reality and across different robot embodiments without retraining, achieving up to 10x reduction in transfer learning tokens. The 'trained model' is a repository of sensorimotor skills, not floating-point weights.
ENPIRE demonstrates fully autonomous physical research with a multi-agent robot fleet
Eight Codex agents were given a fleet of robots, GPUs, and a token budget with the goal of solving tasks as quickly as possible. Without human intervention, they learned to locate visual cues, reset environments, practice novel skills, read online papers, debate approaches, and iterate on hardware. A physical scaling law emerged: more robots in parallel accelerated improvement.
These systems collectively point toward self-improving, continually learning robot ecosystems
RoboTTT's on-the-fly adaptation, ASPIRE's compounding skill library, and ENPIRE's autonomous research loop form a pipeline where robots can learn from experience, share knowledge, and autonomously discover new capabilities. All code will be open-sourced, enabling replication and extension.