Vision foundation models got their ImageNet moment. FTP-1 asks: What if tactile manipulation had its own foundation model? The challenge is that tactile data is far messier than images. A GelSight image, a fingertip pressure array, and a force-torque sensor all produce completely different signals. FTP-1 is the first attempt to build a generalist tactile policy that works across them all. It is pretrained on ~3,000 hours of tactile manipulation data from 26 datasets spanning 21 different tactile sensors.

The most interesting result isn't higher performance on known robots. It's transfer. After pretraining, FTP-1 can adapt to previously unseen tactile sensors while reusing the learned tactile knowledge. Across experiments: • 17.2% success on seen sensor setups • 31% success on unseen tactile-sensor setups This feels like an important shift: Robotics may be moving from "build a policy for a sensor" → "pretrain tactile intelligence once, then plug in new sensors later." For contact-rich manipulation, FTP-1 could become the tactile equivalent of what VLM pretraining did for robot vision.

Most multi-robot VLA systems either centralize everything or train a separate policy per robot. CHORUS asks a different question: can a single pretrained VLA coordinate an entire team using only each robot's local observations? The answer is yes. Each robot runs the same policy, receives a robot-specific identity prompt, and collaborates without communication, shared state, or per-robot policies.

Results are surprisingly strong. Across real-world tasks including tape measurement, book handovers, laundry basket lifting, and 3-robot transport, CHORUS improves success rates by 64 percentage points over decentralized policies trained from scratch, boosts teammate reactivity by 40 percentage points, and even outperforms centralized VLA baselines. A strong example of pretrained VLAs enabling scalable robot teamwork.

Generalist robot policies already contain many useful behaviors, but the challenge is invoking the right one for a new task. Flow Reversal Steering (FRS) introduces a simple idea: take a coarse action from a human or VLM, run it backward through a flow-matching policy to recover the latent noise, then denoise it back into a precise in-distribution robot action. Instead of directly executing rough guidance, FRS projects it onto behaviors the policy already knows.

FRS is not just an inference trick. The discovered latent noises can be distilled with Diffusion Steering via Behavioral Cloning (DSBC), training lightweight steering policies in under a minute, or used to bootstrap RL with semantic guidance. On real DROID tasks, DSBC significantly outperforms standard behavioral cloning, while FRS-bootstrapped RL succeeds on tasks where conventional RL makes little progress.

Robotic visual navigation under partial observability requires foresight: anticipating how moving changes your view and if it brings you closer to the goal. Prior world models predict these futures but rely on slow, external CEM-style planners to choose actions.

By unifying prediction and control, NavWAM acts as a closed-loop policy out of the box. In evaluations, it outperforms planning-based world models without needing test-time action search, matches a much larger 7B VLA policy, and transfers successfully to real mobile robots.