Axis Weekly
Last week, we improved policy takeover and verification, expanded TaskGen's articulated-scene coverage, and ran a second DAgger round.
An eight-step policy-takeover constraint and cleaner derived tasks now sit alongside domain randomization, published LIBERO Pro, and DreamZero finetuning on Axis tasks.
Key updates:
• Takeover & verification: Policy takeover is capped at 8 steps, undo/redo no longer corrupts replay, gripper sensitivity is retuned, and the verification + Human-Gated DAgger (HG-DAgger) pipeline is more automated.
• Derived tasks & articulated coverage: TaskGen now derives clean single-scene tasks and auto-resizes grasped objects. A broader set of articulated tasks (appliances, drawers, cabinets) is live.
• DAgger Round 2 & DreamZero: Round-two data lifted success from 68/160 to 78.3/160 across 3 seeds (+10.3), with HG-DAgger candidates gaining 4-7 points vs baseline. DreamZero is connected for finetuning.
• Randomization & sim-to-real: Domain randomization is in place, LIBERO Pro is published, Isaac Sim rendering is faster, and a sim-to-real hardware setup is ready.
Details below 👇
Axis Weekly
Last week, we closed the remaining replay and runtime gaps between the browser, policy server, and physics stack — then scaled the coverage of our task generation engine (TaskGen).
An articulated asset library is now in the generation path, the full RoboCasa scene grid is online, and a cleaner long- versus short-horizon task split is ready for training and distillation.
Key updates:
• Replay & runtime: Policy and human-trajectory replay now match the frontend path, with web runtime reaching full replay fidelity. Scene loads faster, and simulation pauses automatically when idle.
• Articulated assets & RoboCasa: TaskGen now includes a library of 27 articulated object families (4 variants each). All RoboCasa scenes are online as a 50×50 layout and style grid.
• Horizon splits & next DAgger step: A cleaner long- vs. short-horizon split is ready, LIBERO Pro shows consistently high success rates, and the next round focuses on filtering correction segments that change closed-loop behavior.
Details below 👇