Skip to content

DVAC on Pi05/YAM Reproduction Record

1. Scope and Claim Boundary

  • Method: DVAC, Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot Policies.
  • Paper: arXiv:2606.03847v1.
  • Paper source: arXiv source archive retrieved on 2026-08-24.
  • Official repository: no author implementation was located on 2026-08-24.
  • Official primary backbone: a Pi0.5-based flow policy.
  • ManiMux backbone: the existing JAX Pi05/YAM step-1000 checkpoint and its matching norm stats.
  • No-retraining claim: model parameters, checkpoint, inputs, output transforms and ten-step Euler flow integration are unchanged.
  • Current gate: targeted ManiMux/XPolicy unit tests, real Pi05 GPU forward, WebSocket metadata, two stateful rolling-window probes and one dual-YAM hardware rollout passed.

This integration is a paper reproduction, not an official-code port. The equations, execution order and published defaults are preserved. Two details omitted by the paper are implemented as explicit, reviewable conventions rather than being presented as author code:

  1. an empty first-call rolling buffer is bootstrapped from that call's current variance sequence;
  2. variance is summed over YAM's 14 valid normalized action dimensions, excluding OpenPI's padded dimensions.

2. Paper Algorithm

Pi05 integrates a velocity field from noise time t=1 to clean time t=0. At Euler step i, DVAC reuses the already computed velocity and forms the clean-action estimate

z_i = x_i - t_i * v_theta(x_i, t_i, s)

For future action index k, action dimension d, and the final L denoising estimates, the paper defines

mean_z[k,d] = (1/L) * sum_i z_i[k,d]
V_s(k) = sum_d (1/L) * sum_i (z_i[k,d] - mean_z[k,d])^2

The rolling calibration window contains every per-step variance from up to m recent policy calls:

W_s = union_q {V_q(0), ..., V_q(H-1)}
tau_s = mean(W_s) + alpha * std(W_s)

The first index whose variance is strictly greater than tau_s terminates the stable prefix. If no index crosses the threshold, N_max actions execute. ManiMux uses the bounded operational form

N_exec = N_max                              if no crossing
N_exec = max(N_min, first_crossing)         otherwise

This intentionally follows Equation 7 exactly: despite its name, N_max appears only in the no-crossing branch. The distinction is inert in the current experiment because N_max = H = 50.

Published defaults used here:

Parameter Value Meaning
L 5 final denoising clean estimates
alpha 2.0 rolling standard-deviation tolerance
m 5 recent policy calls in the rolling window
N_min 1 minimum executed prefix
N_max 50 current YAM action horizon

The method performs the same ten velocity evaluations as ordinary Pi05. It does not add VJP, candidate sampling, inverse flow or an auxiliary network.

3. Paper Omissions and Declared Conventions

3.1 Empty rolling buffer

Algorithm 1 computes the threshold from the prior rolling window and appends the current variance after selection, but it does not define the first request when the prior window is empty. This reproduction uses the current variance sequence for the first threshold only, then appends it to the ordinary history. Metadata reports:

cold_start: true
cold_start_policy: current_variance_bootstrap

Every later request calibrates from prior calls only, matching the order in Algorithm 1. XPolicy clears this history on the protocol reset, so statistics never leak across episodes.

3.2 Valid action dimensions

The paper writes D as the action dimension. The current OpenPI model tensor is padded to 32D, while the YAM dataset, norm stats and returned action contain exactly 14 values:

[left 6 joints, left gripper, right 6 joints, right gripper]

DVAC therefore computes the equation over the first 14 normalized model-action dimensions. It does not include the 18 padded dimensions and does not mix physical joint units before summation. The ordinary output transform still unnormalizes, restores absolute joint actions and trims padding exactly as before.

4. Ownership Boundaries

OpenPI sampler

openpi/models/pi0.py adds an isolated return_denoising_variance branch. During each existing Euler step it stores z_i in a fixed five-entry rolling tail. Ordinary, RTC, AAC, PAINT and AutoHorizon branches do not request or allocate this output.

openpi/policies/policy.py slices the clean estimates to the declared valid action dimension and computes Equation 4 before the normal action output transform. It returns the full action chunk plus the variance sequence; it does not select or truncate robot commands.

XPolicy Pi05 adapter

XPolicyLab/policy/Pi_05/model.py owns the episode-local rolling variance history, threshold and prefix selector. get_action_dvac returns the full decoded 50 x 14 YAM chunk plus metadata.

XPolicy WebSocket

The server advertises dvac only when the loaded model exposes get_action_dvac. One request is:

sampling:
  mode: dvac
  tail_steps: 5
  alpha: 2.0
  rolling_window_size: 5
  min_execution_steps: 1
  max_execution_steps: 50

ManiMux Runtime

manimux/runtime/dvac.py follows the paper's synchronous cadence:

  1. wait until the selected prefix is exhausted;
  2. request a new full Pi05 chunk from the latest observation;
  3. validate dvac.execution_steps;
  4. retain exactly that prefix;
  5. execute it through the unchanged Timeline, Executor, Safety and RobotBase;
  6. request again after exhaustion.

This is not RTC or PAINT. Inference is not overlapped with execution, so a physical robot can hold after a short selected prefix while the next request runs. blend_policy_steps is fixed to zero so Timeline does not rewrite the paper-selected prefix.

5. Configuration

server: manimux/configs/policy/pi05/yam/finetune-pick-red-ball-box-step1000.yaml
infra:  manimux/configs/experiments/pick_red_object/pi05/yam_pi05_dvac_step1000.yaml

The server, checkpoint and norm stats are identical to ordinary Pi05. Only the infra runtime and DVAC method parameters change.

6. User-Run Validation

Run the hardware-free contract tests first:

cd /home/ubuntu/manimux
envs/yam/.venv/bin/python -m pytest -q \
  tests/unit/test_dvac_runtime.py \
  tests/unit/test_xpolicylab_plugins.py

PYTHONPATH=/home/ubuntu/manimux/XPolicyLab/policy/Pi_05/openpi/src:/home/ubuntu/manimux \
  XPolicyLab/policy/Pi_05/openpi/.venv/bin/python -m pytest -q \
  XPolicyLab/tests/unit/test_pi05_dvac_variance.py \
  XPolicyLab/tests/unit/test_pi05_dvac_adapter.py \
  XPolicyLab/tests/unit/test_ws_infer_sampling.py

Start the unchanged Pi05 server:

cd /home/ubuntu/manimux
XPolicyLab/policy/Pi_05/openpi/.venv/bin/python \
  manimux/servers/pi05.py \
  --config manimux/configs/policy/pi05/yam/finetune-pick-red-ball-box-step1000.yaml

Before any camera, CAN or robot process, run one hardware-free forward probe:

cd /home/ubuntu/manimux
envs/yam/.venv/bin/python scripts/validation/xpolicylab_yam_forward_probe.py \
  --config manimux/configs/experiments/pick_red_object/pi05/yam_pi05_dvac_step1000.yaml

Required output:

  • finite native and canonical shape 50 x 14;
  • dvac.variance contains 50 finite non-negative scalars;
  • 1 <= dvac.execution_steps <= 50;
  • tail_steps=5, action_dim=14, variance_space=normalized_valid_action;
  • first request reports cold_start=true.

Then exercise the rolling window with three requests in one protocol session:

cd /home/ubuntu/manimux
envs/yam/.venv/bin/python scripts/validation/xpolicylab_yam_dvac_probe.py \
  --config manimux/configs/experiments/pick_red_object/pi05/yam_pi05_dvac_step1000.yaml \
  --requests 3

Required output:

  • cold_start is [true, false, false];
  • rolling_states is [1, 2, 3];
  • every request returns a finite full 50 x 14 chunk before Runtime truncation;
  • every selected execution_steps is in [1, 50].

After the normal YAM camera, CAN, achieved-state, start-pose and emergency-stop checks, only the operator may run:

cd /home/ubuntu/manimux
envs/yam/.venv/bin/manimux run \
  --config manimux/configs/experiments/pick_red_object/pi05/yam_pi05_dvac_step1000.yaml

DVAC may select a very short prefix. The unchanged synchronous paper cadence then holds the last command during the next model request; this is expected behavior, not a missing chunk.

7. Validation Matrix

Gate Status Evidence required
Paper/source audit complete arXiv v1 equations, algorithm and defaults
Official code parity unavailable no author repository located
Targeted unit contracts passed ManiMux 39 passed; XPolicy current checkout 14 passed
OpenPI sampler contract GPU exercised finite variance from L=5, valid action_dim=14
Pi05 adapter/rolling state passed by probe cold start then prior-history states 1,2,3
WebSocket capability/dispatch passed by probe three DVAC replies in one session
ManiMux synchronous truncation unit passed full decode then prefix selection
Pi05 JAX GPU forward passed finite 50 x 14, warmed 96.4/88.9 ms
YAM hardware exercised operator completed one dual-YAM rollout
Task benefit unverified controlled repeated trials

8. Reviewer Checklist

  • [ ] Ordinary Pi05 output is unchanged when DVAC is not requested.
  • [ ] Exactly the final five clean estimates are used.
  • [ ] z_i = x_i - t_i v_i uses OpenPI's noise-to-clean time convention.
  • [ ] Variance uses divisor L, sums valid action dimensions and excludes padding.
  • [ ] Threshold uses prior calls, except the disclosed first-call bootstrap.
  • [ ] Current variance is appended only after selection.
  • [ ] Strict crossing is V_s(k) > tau_s, not >=.
  • [ ] XPolicy returns a full chunk; ManiMux alone truncates it.
  • [ ] Protocol reset clears the rolling buffer.
  • [ ] blend_policy_steps=0; no RTC, AAC, PAINT or AutoHorizon mode is composed implicitly.
  • [ ] Synchronous inference holds are reported rather than mislabeled as missing chunks.
  • [ ] Unit, GPU, WebSocket and hardware evidence remain separate.

9. First Stateful GPU Evidence

On 2026-08-24 the operator ran three DVAC requests against the real local Pi05 step-1000 checkpoint without cameras, CAN or robot motion. All requests returned finite full 50 x 14 actions and normalized 14D variance metadata. The first request compiled JAX in 5292.6 ms; warmed requests took 96.4 ms and 88.9 ms.

The rolling state advanced exactly as required:

request       1      2      3
cold_start  true   false  false
history        1      2      3
N_exec        50     50     46
crossing    none   none     46

This validates the real sampler, protocol envelope and episode-local rolling selector. It does not validate ManiMux physical execution, task success or a performance improvement over another method.

An independent current-checkout probe then returned warmed latencies 107.0/94.4/93.1 ms and selected horizons 47/50/3. This confirms that the adaptive selector is active rather than silently falling back to a fixed horizon. It also predicts that stable phases may execute long continuous prefixes, while a three-step uncertain prefix can expose an approximately one-inference-duration hold before the next chunk arrives.

10. First YAM Hardware Evidence

On 2026-08-24 the operator ran the DVAC configuration on dual YAM with the local Pi05 step-1000 checkpoint. The robot executed model output successfully. Motion remained visibly stop-and-go, like a serial inference loop, but the operator judged the physical trajectory substantially more accurate than the immediately preceding adaptive-horizon method.

This observation is consistent with DVAC discarding a future suffix after denoising variance crosses the rolling threshold and replanning from a newer observation. It is one qualitative rollout, not a controlled success-rate result, and does not prove that DVAC is generally more accurate. The pauses remain expected because the faithful Runtime waits for the selected prefix to finish before issuing the next synchronous model request.