Imitation Learning

π0.5 is the default baseline, and papers target its failures

Physical Intelligence's π0.5 appears as the base policy or System 1 controller in a long list of recent papers, including DROID real-world experiments, harness ablations and DreamTrue's offline policy evaluation. It is also the comparison point that new methods beat. Long-WAM reaches 90-100% grasping on a Unitree G1 where π0.5 falls from 15% to 0%. Another paper shows π0.5 failing 4/5 coat trials with higher prediction delay than GR00T. Multi-subgoal studies find generalist policies collapsing from 85.4% single-subgoal success to 0% at five subgoals in simulation. The model's authors (Levine, Finn, Black) are cited as the reference lineage of the field.

Researchers are questioning core design choices of imitation-learned VLAs. One paper finds discrete action parametrizations match or beat continuous ones across RoboMimic, RoboCasa and LIBERO, directly challenging the continuous flow design of π0 and π0.5. SkipVLA routes between a VLA and a classical motion planner for a 2.5x speedup and over 52% energy savings on Jetson Thor. ARC shows external VLM reasoning can lift zero-shot performance without new robot data. Interpretability work reports that VLA backbones condition the action expert only through a narrow middle-to-late region.

Why it matters · If cheaper, hybrid or discrete designs match the flagship recipe, the moat shifts from the model architecture to data, evaluation and deployment efficiency.

Work is moving beyond cloning demonstrations toward whole-body compliance, environment adaptation and trustworthy evaluation. C-GMT adds a residual compliance adapter to a frozen motion tracker, and full-body compliance is shown to beat upper-body-only control. Splat2Act personalizes a VLA from a phone scan, with 66.7% simulation and 63.9% real-robot success. World models such as Ropedia W0 rank π0.5 checkpoints with r=0.984 correlation without physical access. Value caching targets soft aliasing in robot RL.

Why it matters · Evaluation and adaptation tooling that works without physical robots is becoming a distinct product category and lowers the cost of iterating on imitation policies.

Demonstration data remains the bottleneck, and the ecosystem is widening it. Ego4D's roughly 4,000 hours of human demonstration was the reference point that motivated Mecka's data-collection work. AgiBot World supplies public pretraining data for Long-WAM (4.8% of its corpus), and BitRobot and Lightwheel's EgoSuite are emerging as new datasets. DreamTrue finds dataset camera calibration frequently wrong, improving alignment on 97.1% of AgiBot episodes.

Why it matters · Data quality, calibration and provenance are becoming investable infrastructure layers beneath the model companies.

Weekly funding is $0M in most weeks, with spikes of $117M across 4 deals in late August and $72M across 2 deals in the week of Oct 12. The Oct 12 week is dominated by a $36M Series B led by Ascensia, reported by both PitchBook and StrictlyVC. Mentions run at 24-27 per week in early October, almost entirely arXiv papers. Deal flow lags far behind research attention.

Why it matters · Investors must pick from a research-heavy field with few priced rounds, so early access to the labs and data players matters more than chasing announced deals.