Curated digests of new papers and releases in embodied AI and robot learning, with what each result does and does not show.
ActionUNet adds a temporal U-Net and SIREN decoder to a frozen VLA; authors report π0.5 gains of 9.8/6.1/11.4 points in robustness tests, 6.46% slower calls.
Preprint: 4-bit formats on Orin NX differ 36.8× in step energy; IQ4_NL has the lowest energy and resident memory, while NF4 and INT4 also run on-board.
Frozen VLA plus a sim-only residual on VLM latents aligned by an optimal-transport mapper; authors report real-robot gains in all 8 task–backbone pairs.
π0.5-derived waypoint VLA: plans took 57 serial passes, endpoint erasure didn't hurt success. Authors cut robot planning 1094→125 ms; NGM makes plans count.
Sequential robot policy comparison via safe anytime-valid inference, for non-binary metrics too; authors report up to 70% less evaluation vs batch tests.
14 items from the Physical AI Radar's September archive, numbers checked: VLA evaluation starts testing recovery; latency becomes a formally reported metric.
7 primary sources: recovery benchmarks, sequential testing, post-quantization closed-loop checks, sim fidelity. 98.83% initial success, only 48% recovery.
Half-day SFT, RL post-training, learning from deployment: 7 radar sources outline VLA post-training, and the risk that online RL silently breaks single tasks.
9 sources: Rolling-WAM cuts replanning from 978 to 215 ms, early exits plus fewer denoising steps cut ~80% of latency; quantization saves memory, not time.
Sign in to chat with the assistant.
Sign in