Anonymous submission · under double-blind review · ICLR 2027
The 6 Kinetix levels whose state includes objects beyond proprioception, which we treat as the slow channel. H = 8, 2,048 rollouts per level, 3 seeds.
| Method | d₀ = 1 | d₀ = 2 | d₀ = 3 | d₀ = 4 |
|---|---|---|---|---|
| Standard Flow (async) | 0.655 ± 0.000 | 0.569 ± 0.001 | – | – |
| RTC | 0.838 ± 0.000 | 0.690 ± 0.003 | – | – |
| Train-time RTC | 0.852 ± 0.017 | 0.794 ± 0.006 | – | – |
| FASTER | 0.826 ± 0.073 | 0.763 ± 0.041 | – | – |
| SDP | 0.180 ± 0.010 | 0.153 ± 0.024 | 0.088 ± 0.003 | 0.091 ± 0.001 |
| πR² (ours) | 0.819 ± 0.004 | 0.804 ± 0.006 | 0.769 ± 0.008 | 0.769 ± 0.008 |
| πR² w/o async slow channel | 0.822 ± 0.004 | 0.807 ± 0.010 | 0.744 ± 0.029 | 0.646 ± 0.017 |
| πR² w/o staircase (= RTC + async VLM) | 0.812 ± 0.017 | 0.776 ± 0.021 | 0.707 ± 0.027 | 0.642 ± 0.059 |
| πR² w/o latency-adaptive | 0.210 ± 0.031 | 0.190 ± 0.035 | 0.197 ± 0.030 | 0.197 ± 0.030 |
Each method is charged its own deployment delay (R4); “–” means that delay no longer fits the chunk (H/2 = 4). πR² and its ablations receive the measured cache age dvis through the delay embedding, as on the robot; the baselines have no cache. Bold: methods at or above 0.80, i.e. comparable at that delay; where none reaches it, the best.
Left: overhead camera. Right: d, the policy call interval, and dvis, the age of the cached VLM embedding — both re-measured every call. Dashed line: the 5-tick training range.
dvis varies continuously around a median of ~2.5 ticks; in five of the twelve rollouts it leaves the 5-tick training range, peaking at 13.9 ticks (2.8× the range) in one. The call interval holds within ±10% on 93–100% of calls and spikes to at most 7.0 ticks when a chunk arrives late. All twelve rollouts succeed.
Measured on the robot’s compute, 1 tick = 40 ms. πR² on one GPU still beats every RTC configuration, including RTC with an asynchronous VLM on two.
| Method | GPUs | VLM | Policy latency d | VLM latency dvlm | Cache age dvis | Proprio staleness spro | VLM staleness svis | svis > 5 ticks |
|---|---|---|---|---|---|---|---|---|
| πR² | 2 | async | 31.5 ms (0.79) | 60.8 ms (1.52) | 94 ms (2.36) | 48 ms (1.21) | 141 ms (3.52) | 0.0% |
| πR² | 1 | async | 69.3 ms (1.73) | 69.4 ms (1.73) | 88 ms (2.21) | 104 ms (2.61) | 194 ms (4.85) | 40.5% |
| Train-time RTC | 1 | sync | 133.3 ms (3.33) | – | – | 203 ms (5.08) | 203 ms (5.08) | 50.8% |
| Train-time RTC | 2 | async | 94.4 ms (2.36) | 59.8 ms (1.49) | 96 ms (2.40) | 142 ms (3.55) | 235 ms (5.87) | 83.8% |
Ticks in parentheses. Staleness is the median over executed control ticks and counts each action’s offset in its chunk. “–”: already included in d.
In simulation the policy latency and cache age are set by a fixed cost model from GR00T-N1.7 — VLM latency dvlm = 0.75d₀ and denoising-step latency ddn = 0.25d₀, where the unit delay d₀ is the cost of GR00T’s default 4 denoising steps — not by how many denoising steps the simulation network runs. Flow and RTC run 15 steps on Leap Cube and up to 5 on Kinetix for quality, but are charged only GR00T’s 4, which favors them. A method is excluded at a given d₀ when its charged policy latency exceeds H/2 (H = 16 on Leap Cube, H = 8 on Kinetix), as it then no longer allows continuous real-time execution.
| Method | Charged policy latency d | d | dvis |
|---|---|---|---|
| Standard Flow, RTC, Train-time RTC | dvlm + 4ddn = ⌈1.75d₀⌉ | 2 / 4 / 6 / – | – |
| FASTER | ⌈0.75d₀ + (n(d)/K)·d₀⌉ | 1 / 3 / 4 / – | – |
| SDP | ⌈d₀⌉, stalls | 1 / 2 / 3 / 5 | – |
| πR² | 1·ddn = ⌈0.25d₀⌉ | 1 / 1 / 1 / 2 | 1 / 2 / 3 / 4 |
| πR² w/o async | dvlm + 1·ddn = ⌈1.0d₀⌉ | 1 / 2 / 3 / 5 | – |
| πR² w/o staircase | 4·ddn = ⌈1.0d₀⌉ | 1 / 2 / 3 / 5 | 1 / 2 / 3 / 4 |
| πR² w/o latency-adaptive | ⌈0.25d₀⌉, stalls | 1 / 1 / 1 / 2 | 1 / 2 / 3 / 4 |
| Method | Charged policy latency d | d | dvis |
|---|---|---|---|
| Standard Flow, RTC, Train-time RTC | dvlm + 4ddn = ⌈1.75d₀⌉ | 2 / 4 / – / – | – |
| FASTER | ⌈0.75d₀ + (n(d)/K)·d₀⌉ | 1 / 4 / – / – | – |
| SDP | ⌈d₀⌉, stalls | 1 / 2 / 3 / 4 | – |
| πR² | 1·ddn = ⌈0.25d₀⌉ | 1 / 1 / 1 / 1 | 1 / 2 / 3 / 3 |
| πR² w/o async | dvlm + 1·ddn = ⌈1.0d₀⌉ | 1 / 2 / 3 / 4 | – |
| πR² w/o staircase | 4·ddn = ⌈1.0d₀⌉ | 1 / 2 / 3 / 4 | 1 / 2 / 3 / 3 |
| πR² w/o latency-adaptive | ⌈0.25d₀⌉, stalls | 1 / 1 / 1 / 1 | 1 / 2 / 3 / 3 |
All values in control ticks. “–” in d means the charged latency exceeds H/2, so the schedule cannot fit a clean prefix and a noise tail of length d inside the chunk; in dvis it means the VLM latency is already inside d. For FASTER, n(d) is the ready-step of the last position it emits under its own schedule (α = 0.6, u₀ = 0.9), with K = 15 steps on Leap Cube and K = 5 on Kinetix; because the schedule depends on the delay, d is the smallest self-consistent solution. πR²’s cache age is charged as dvis = ⌈0.75d₀⌉, the lower end of its real-world range.