OpenTau 0.14 adds Xiaomi-Robotics-1 and fixes RoboCasa action order

OpenTau 0.14 ports Xiaomi-Robotics-1, adds best-of-N action sampling, and brings an experimental π0.5 with test-time-training memory. It also fixes a RoboCasa bug that scrambled every action, and it renames the gRPC key, which matters before you upgrade a server.

Slide text

OpenTau

v0.14.0

Release notes

• Xiaomi-Robotics-1
• Best-of-N sampling
• pi05_ttt memory
• RoboCasa actions
• Other fixes
• gRPC auth rename
• Upgrading

OpenTau 0.14.0

OpenTau 0.14 adds Xiaomi-Robotics-1 and fixes RoboCasa action order

xr1 policy best-of-N sampling pi05_ttt memory RoboCasa actions fixed gRPC auth renamed

The biggest addition is a new policy type, xr1, a port of Xiaomi-Robotics-1. A Qwen3-VL-4B vision-language model fills a key-value cache at each of its 36 layers, and a 36-layer diffusion transformer reads that cache, layer by layer, to produce the action chunk by flow matching.

Against the reference checkpoint the port is bit-identical, from the prompt tokens to the decoded action chunk. In RoboCasa simulation its success rates land within the confidence intervals of the published ones. To reproduce them, start from the shipped eval config, which puts the wrist camera last and loads both object registries.

Best-of-N sampling denoises several noise draws per observation and lets a critic keep one chunk. The vision-language prefix runs once, so on an RTX 5090 eight candidates take about 156 milliseconds, against 153 for one. Replicating the observation instead takes about 390. It is off by default, covers pi05 and pi06 only, and its one critic picks the consensus chunk, which is not a judge of quality.

pi05_ttt is a new, experimental policy that ports RoboTTT onto π0.5. A test-time-training layer follows attention in each of the action expert's 18 layers, and its fast weights update at every timestep, so memory carries across steps. Around it come paired demonstration and rollout episodes, trainable-only snapshots, and inference diagnostics.

If you evaluate in RoboCasa, this fix changes your numbers. OpenTau's in-process environment read the 12-number action with base motion first. RoboCasa puts the end effector first, so the arm command drove the base. Episodes still ran, and only the success rate fell. Treat every earlier RoboCasa number as understated, and re-run your baselines.

A few more fixes. Older π0.5 checkpoints with pre-rename normalizer keys load again. Multi-rank LIBERO eval across all tasks no longer crashes. The gRPC server works after torch compile in max-autotune mode. And the value policy reads a learned readout token, instead of padding that limited it to the mean return.

One change breaks existing setups. The gRPC server's API key moved: clients send x-api-key, and the server reads INFERENCE_API_KEY. The old names are gone. A client on the old header fails closed, with an error naming the right one. A launcher still exporting the old variable fails open, and the server runs unauthenticated without a warning.

To upgrade, run pip install with the upgrade flag. When you do, rename the gRPC key on servers that use auth, drop sequence stride from sequence configs, and re-run RoboCasa baselines. If a pinned older OpenTau reads checkpoints you save, upgrade it too, since their configs now carry two keys it rejects. That is OpenTau 0.14: Xiaomi-Robotics-1, best-of-N, and RoboCasa actions in the right order.