cognaclee's picture
Add files using upload-large-folder tool
cd30adb verified
Raw History Blame Contribute Delete
25 kB
"""AdaSpeVLA plugin on the OpenPI pi0.5 backbone.
The PaliGemma prefix (images + tokenized prompt with discrete state) is exactly
pi0.5's; the pi0.5 action expert is removed and replaced with the AdaSpeVLA heads
ported from the v1.3 PyTorch reference (adaspevla/models/{vlm_backbone,
bspline_flow_policy,adaspe_vla_policy}.py, bspline/learnable_knots.py):
prefix hidden states -> fusion (pooled cond + token memory)
-> learnable knot head -> B-spline flow-matching transformer -> control points
-> timing (query-interval) head.
`generator` picks what produces the control points:
"flow_head" (default; the 2026-11-12 runs): the 20 M AdaSpeVLA flow transformer above, trained from scratch,
cross-attending to the projected last-layer prefix tokens.
"pi05_expert" (2026-11-13): pi0.5's pretrained Gemma-300M action expert (weights from pi05_base) runs on K
control-point tokens exactly as pi0.5 runs on its action tokens: per-layer attention to the prefix
KV cache, adaRMS time conditioning, pi0.5's flow convention (t=1 noise, Beta(1.5, 1) time) and
10 ODE steps. Each token also gets a zero-initialized projection of where its control point sits on
the predicted knot vector. Knot head, timing head, labels and spline fitting are unchanged.
The knots come from the prefix, so the expert runs after the prefix pass on its KV cache (as pi0.5 does at
inference); gradients flow through the cache, so training equals pi0.5's joint prefix+suffix pass.
Actions are OpenPI-normalized 7D body-frame deltas over `action_horizon` steps.
"""
import dataclasses
import math
from typing import Literal
import einops
import flax.nnx as nnx
import jax
import jax.numpy as jnp
from typing_extensions import override
from openpi.models import gemma as _gemma
from openpi.models import model as _model
from openpi.models import pi0_config
from openpi.models import siglip as _siglip
from openpi.models.pi0 import make_attn_mask, posemb_sincos
import openpi.shared.array_typing as at
import flax.nnx.bridge as nnx_bridge
from . import bspline as bs
from . import labels
from .labels import TIMING_CHANNELS # noqa: F401 (re-exported for adaspe/timing_data.py)
CP_POS_DIM = 64 # sin-cos size of a control point's knot position (pi05_expert tokens)
@dataclasses.dataclass(frozen=True)
class AdaSpePi05Config(pi0_config.Pi0Config):
"""pi0.5 config whose action expert is replaced by the AdaSpeVLA plugin."""
pi05: bool = True
stage: Literal["action", "timing"] = "action"
generator: Literal["flow_head", "pi05_expert"] = "flow_head" # see the module docstring
plugin_action_dim: int = 7 # real action channels; the rest of `action_dim` is OpenPI padding
views: tuple[str, ...] = ("base_0_rgb", "left_wrist_0_rgb")
control_dt: float = 0.1
cond_dim: int = 512
num_control_points: int = 8
degree: int = 3
min_interval_eps: float = 0.01
min_query_interval: float = 0.2
max_query_interval: float = 0.8
flow_hidden_dim: int = 384
flow_num_layers: int = 6
flow_num_heads: int = 6
knot_hidden_dim: int = 256
knot_loss_coeff: float = 0.5
flow_loss_coeff: float = 1.0
density_gain: float = 10.0
gripper_density_gain: float = 10.0
gripper_threshold: float = 0.5
e_crit: float = 1.0 # set from training-data calibration
# Timing labels. "command_delta": events on the normalized per-step commands (before 2026-11-09).
# "measured": as AdaSpeVLA franka v1.x, events on measured anchor-relative pose offsets divided by
# per-channel motion scales, and on the measured gripper width crossing gripper_threshold. The
# measured window (state[4:10] pose, state[0] width; 13 frames) arrives in action channels
# TIMING_CHANNELS, which are zero padding for the plugin (adaspe/timing_data.py).
# "measured_cmd_gripper": measured motion as "measured", gripper events from the commanded open/closed
# state instead of the measured width, which lags the command by ~1.9 s on the franka data (adaspe/labels.py).
timing_label_mode: Literal["command_delta", "measured", "measured_cmd_gripper"] = "command_delta"
# Knot-density supervision, same modes ("measured" = AdaSpeVLA original). The flow target
# stays the delta commands either way; only the knot-density target changes.
knot_label_mode: Literal["command_delta", "measured", "measured_cmd_gripper"] = "command_delta"
# "e_crit": knot density uses tanh(motion_variation / e_crit) instead of the saturating tanh(motion_variation).
knot_motion_norm: Literal["none", "e_crit"] = "none"
timing_motion_scale: tuple[float, ...] | None = None # 95th pct |offset| per channel (calibrated; knots too)
timing_probe_count: int = 16
timing_num_ode_steps: int = 15
num_ode_steps: int = 15
@override
def create(self, rng: at.KeyArrayLike) -> "AdaSpePi05":
return AdaSpePi05(self, rngs=nnx.Rngs(rng))
@property
def horizon_duration(self) -> float:
return (self.action_horizon - 1) * self.control_dt
def get_freeze_filter(self) -> nnx.filterlib.Filter:
if self.stage == "timing":
return nnx.All(nnx.Param, nnx.Not(nnx_utils_path(".*freq_head.*")))
return nnx.Nothing
def nnx_utils_path(regex):
from openpi.shared import nnx_utils
return nnx_utils.PathRegex(regex)
class Seq(nnx.Module):
def __init__(self, *layers):
self.layers = list(layers)
def __call__(self, x):
for layer in self.layers:
x = layer(x)
return x
# Activations are module-level classes: a class defined per call would give the model a different
# graph type on every build, which breaks OpenPI's jit out_shardings check.
class _SiLU(nnx.Module):
def __call__(self, x):
return nnx.silu(x)
class _GELU(nnx.Module):
def __call__(self, x):
return jax.nn.gelu(x, approximate=False)
class _Sigmoid(nnx.Module):
def __call__(self, x):
return jax.nn.sigmoid(x)
def silu():
return _SiLU()
def gelu():
return _GELU()
def sigmoid():
return _Sigmoid()
def ln(dim, rngs):
return nnx.LayerNorm(dim, epsilon=1e-5, rngs=rngs)
def sinusoidal(t, dim):
half = dim // 2
freq = jnp.exp(jnp.arange(half) * -(math.log(10000) / (half - 1)))
arg = t[:, None] * freq[None]
return jnp.concatenate([jnp.sin(arg), jnp.cos(arg)], -1)
class FlowBlock(nnx.Module):
"""AdaLN transformer block with cross-attention to the observation memory."""
def __init__(self, dim, heads, rngs, mlp_ratio=4.0):
self.norm1 = ln(dim, rngs)
self.attn = nnx.MultiHeadAttention(heads, dim, decode=False, rngs=rngs)
self.norm_cross = ln(dim, rngs)
self.cross_attn = nnx.MultiHeadAttention(heads, dim, decode=False, rngs=rngs)
self.norm2 = ln(dim, rngs)
hidden = int(dim * mlp_ratio)
self.mlp = Seq(nnx.Linear(dim, hidden, rngs=rngs), gelu(), nnx.Linear(hidden, dim, rngs=rngs))
self.ada = Seq(silu(), nnx.Linear(dim, 6 * dim, rngs=rngs))
def __call__(self, x, c, memory, memory_mask):
s_msa, sc_msa, g_msa, s_mlp, sc_mlp, g_mlp = jnp.split(self.ada(c)[:, None], 6, axis=-1)
h = self.norm1(x) * (1 + sc_msa) + s_msa
x = x + g_msa * self.attn(h, h, h)
x = x + self.cross_attn(self.norm_cross(x), memory, memory, mask=memory_mask[:, None, None, :])
h = self.norm2(x) * (1 + sc_mlp) + s_mlp
return x + g_mlp * self.mlp(h)
class FlowPolicy(nnx.Module):
def __init__(self, cfg: AdaSpePi05Config, context_dim, rngs):
d, k, a = cfg.flow_hidden_dim, cfg.num_control_points, cfg.plugin_action_dim
self.dim = d
self.input_proj = nnx.Linear(a, d, rngs=rngs)
self.pos_emb = nnx.Param(jax.random.normal(rngs.params(), (1, k, d)) * 0.02)
self.cond_proj = Seq(nnx.Linear(context_dim, d, rngs=rngs), silu(), nnx.Linear(d, d, rngs=rngs))
self.time_mlp = Seq(nnx.Linear(d, d, rngs=rngs), silu(), nnx.Linear(d, d, rngs=rngs))
self.memory_proj = Seq(ln(cfg.cond_dim, rngs), nnx.Linear(cfg.cond_dim, d, rngs=rngs))
self.blocks = [FlowBlock(d, cfg.flow_num_heads, rngs) for _ in range(cfg.flow_num_layers)]
self.final_norm = ln(d, rngs)
self.output_proj = nnx.Linear(d, a, rngs=rngs)
def __call__(self, x_t, t, context, memory, memory_mask):
c = self.time_mlp(sinusoidal(t, self.dim)) + self.cond_proj(context)
h = self.input_proj(x_t) + self.pos_emb.value
mem = self.memory_proj(memory)
for block in self.blocks:
h = block(h, c, mem, memory_mask)
return self.output_proj(self.final_norm(h))
class AdaSpePi05(_model.BaseModel):
def __init__(self, config: AdaSpePi05Config, rngs: nnx.Rngs):
super().__init__(config.action_dim, config.action_horizon, config.max_token_len)
self.cfg = config
pg = _gemma.get_config(config.paligemma_variant)
expert = config.generator == "pi05_expert"
if expert: # pi0.5's two-stream Gemma (Pi0.__init__): PaliGemma + the action expert with adaRMS
ex = _gemma.get_config(config.action_expert_variant)
llm = nnx_bridge.ToNNX(_gemma.Module(configs=[pg, ex], embed_dtype=config.dtype, adarms=True))
llm.lazy_init(rngs=rngs, method="init", use_adarms=[False, True])
else: # PaliGemma only: the pi0.5 action expert is replaced by the plugin flow head.
llm = nnx_bridge.ToNNX(_gemma.Module(configs=[pg], embed_dtype=config.dtype, adarms=True))
llm.lazy_init(rngs=rngs, method="init", use_adarms=[False])
img = nnx_bridge.ToNNX(
_siglip.Module(num_classes=pg.width, variant="So400m/14", pool_type="none", scan=True, dtype_mm=config.dtype)
)
img.lazy_init(next(iter(config.fake_obs().images.values())), train=False, rngs=rngs)
self.PaliGemma = nnx.Dict(llm=llm, img=img)
f, c, v = pg.width, config.cond_dim, len(config.views)
n_knots = config.num_control_points + config.degree + 1
n_spans = config.num_control_points - config.degree
# Observation fusion (vlm_backbone.py, spatial-token mode).
self.fusion_proj = Seq(nnx.Linear(f * v, c, rngs=rngs), ln(c, rngs), silu())
self.text_proj = Seq(nnx.Linear(f, c, rngs=rngs), ln(c, rngs), silu())
self.cross_gate = Seq(nnx.Linear(2 * c, c, rngs=rngs), sigmoid())
self.final_proj = nnx.Linear(c, c, rngs=rngs)
if not expert: # token memory for the flow head's cross-attention (the expert reads the prefix KV cache)
self.token_proj = nnx.Linear(f, c, rngs=rngs)
self.view_embedding = nnx.Param(jax.random.normal(rngs.params(), (v, 1, c)) * 0.02)
self.text_token_proj = nnx.Linear(f, c, rngs=rngs)
# Knot head (learnable_knots.py).
h = config.knot_hidden_dim
self.knot_head = Seq(
nnx.Linear(c + 1, h, rngs=rngs), silu(), nnx.Linear(h, h, rngs=rngs), silu(), nnx.Linear(h, n_spans, rngs=rngs)
)
if expert:
# Same names/shapes as Pi0 (pi05=True), so PluginWeightLoader takes them from pi05_base.
w = ex.width
self.action_in_proj = nnx.Linear(config.action_dim, w, rngs=rngs)
self.time_mlp_in = nnx.Linear(w, w, rngs=rngs)
self.time_mlp_out = nnx.Linear(w, w, rngs=rngs)
self.action_out_proj = nnx.Linear(w, config.action_dim, rngs=rngs)
# New: control-point position on the knot vector (+ knots, log horizon), added to each token.
# Zero init, so at step 0 the expert gets exactly pi0.5's token inputs.
self.cp_knot_proj = nnx.Linear(
CP_POS_DIM + n_knots + 1, w, kernel_init=jax.nn.initializers.zeros, rngs=rngs)
else:
self.flow_policy = FlowPolicy(config, c + n_knots + 1, rngs)
timing_dim = c + 1 + n_spans + config.timing_probe_count * config.plugin_action_dim
self.freq_head = Seq(nnx.Linear(timing_dim, 128, rngs=rngs), silu(), nnx.Linear(128, 1, rngs=rngs))
self.deterministic = True
# ---- observation encoding -------------------------------------------------
def _prefix(self, obs: _model.Observation):
tokens, masks, ar, view_slices = [], [], [], {}
pos = 0
for name in obs.images:
t, _ = self.PaliGemma.img(obs.images[name], train=False)
tokens.append(t)
masks.append(einops.repeat(obs.image_masks[name], "b -> b s", s=t.shape[1]))
ar += [False] * t.shape[1]
view_slices[name] = (pos, pos + t.shape[1])
pos += t.shape[1]
text = self.PaliGemma.llm(obs.tokenized_prompt, method="embed")
tokens.append(text)
masks.append(obs.tokenized_prompt_mask)
ar += [False] * text.shape[1]
tokens = jnp.concatenate(tokens, 1)
mask = jnp.concatenate(masks, 1)
attn = make_attn_mask(mask, jnp.array(ar))
positions = jnp.cumsum(mask, axis=1) - 1
if self._expert: # also keep the KV cache for the expert's suffix pass (Pi0.sample_actions)
(out, _), kv_cache = self.PaliGemma.llm([tokens, None], mask=attn, positions=positions)
else:
(out,), kv_cache = self.PaliGemma.llm([tokens], mask=attn, positions=positions)
return out.astype(jnp.float32), view_slices, pos, kv_cache, mask
@property
def _expert(self) -> bool:
return self.cfg.generator == "pi05_expert"
def encode(self, obs: _model.Observation):
"""Global condition (B,C) and what the generator reads from the observation (`obs_ctx`): token memory
(B,M,C) + memory mask (B,M) for the flow head, or the prefix KV cache + prefix mask (B,P) for the expert."""
hidden, slices, n_img, kv_cache, prefix_mask = self._prefix(obs)
b = hidden.shape[0]
views = jnp.stack([hidden[:, slices[n][0] : slices[n][1]] for n in self.cfg.views], 1) # (B,V,P,F)
valid = obs.tokenized_prompt_mask[..., None].astype(jnp.float32)
text_hidden = hidden[:, n_img:]
text_feat = (text_hidden * valid).sum(1) / jnp.maximum(valid.sum(1), 1.0)
vision = self.fusion_proj(views.mean(2).reshape(b, -1))
text = self.text_proj(text_feat)
gate = self.cross_gate(jnp.concatenate([vision, text], -1))
cond = self.final_proj(gate * vision + (1 - gate) * text)
if self._expert:
return cond, (kv_cache, prefix_mask)
mem = (self.token_proj(views) + self.view_embedding.value).reshape(b, -1, self.cfg.cond_dim)
memory = jnp.concatenate([mem, self.text_token_proj(text_feat)[:, None]], 1)
return cond, (memory, jnp.ones(memory.shape[:2], bool))
# ---- plugin pieces --------------------------------------------------------
def _horizon(self, b):
return jnp.full((b, 1), self.cfg.horizon_duration, jnp.float32)
def predict_knots(self, cond, horizon):
cfg = self.cfg
n = cfg.num_control_points - cfg.degree
raw = jax.nn.softmax(self.knot_head(jnp.concatenate([cond, jnp.log(horizon)], -1)), -1)
delta = cfg.min_interval_eps + (1 - n * cfg.min_interval_eps) * raw
return bs.intervals_to_knots(delta, cfg.degree), delta
def _context(self, cond, knots, horizon):
return jnp.concatenate([cond, knots, jnp.log(horizon)], -1)
def _times(self, b):
t = jnp.arange(self.action_horizon, dtype=jnp.float32) * self.cfg.control_dt
return jnp.broadcast_to(t, (b, self.action_horizon))
def _gripper01(self, actions):
"""Gripper command as an open(1)/closed(0) state (B,T,1); see labels.gripper_open_state."""
return labels.gripper_open_state(actions[..., 6])
def sample_control_points(self, rng, cond, knots, horizon, obs_ctx, num_steps, noise=None):
b = cond.shape[0]
x = jax.random.normal(rng, (b, self.cfg.num_control_points, self.cfg.plugin_action_dim)) if noise is None else noise
if self._expert: # pi0.5 convention: integrate from t=1 (noise) to t=0 (data), as Pi0.sample_actions
dt = -1.0 / num_steps
bias = self._cp_token_bias(knots, horizon)
def step(i, x):
t = jnp.full((b,), 1.0 + i * dt, jnp.float32)
return x + dt * self._expert_velocity(x, t, bias, obs_ctx)
return jax.lax.fori_loop(0, num_steps, step, x)
context = self._context(cond, knots, horizon)
memory, mask = obs_ctx
dt = 1.0 / num_steps
def step(i, x):
t = jnp.full((b,), i * dt, jnp.float32)
return x + dt * self.flow_policy(x, t, context, memory, mask)
return jax.lax.fori_loop(0, num_steps, step, x)
# ---- pi05_expert generator ----------------------------------------------------
def _cp_token_bias(self, knots, horizon):
"""(B,K,W) added to the control-point tokens: sin-cos of each control point's Greville abscissa
(mean of knots i+1..i+degree, in [0, 1]) plus the knot vector and log horizon, via the zero-init cp_knot_proj."""
cfg = self.cfg
b, k, d = knots.shape[0], cfg.num_control_points, cfg.degree
greville = knots[:, jnp.arange(k)[:, None] + jnp.arange(1, d + 1)[None]].mean(-1) # (B,K)
pos = posemb_sincos(greville.reshape(-1), CP_POS_DIM, min_period=4e-3, max_period=4.0).reshape(b, k, -1)
glob = jnp.concatenate([knots, jnp.log(horizon)], -1)
feats = jnp.concatenate([pos, jnp.broadcast_to(glob[:, None], (b, k, glob.shape[-1]))], -1)
return self.cp_knot_proj(feats)
def _expert_velocity(self, x_t, time, token_bias, obs_ctx):
"""pi0.5's suffix pass on K control-point tokens: Pi0.embed_suffix (pi05=True) + the KV-cache llm call
of Pi0.sample_actions. x_t (B,K,plugin_action_dim) -> velocity, same shape. The OpenPI padding channels
are fed as zeros (pi0.5 pads actions to action_dim) and their outputs are dropped."""
kv_cache, prefix_mask = obs_ctx
b, k, a = x_t.shape
tokens = self.action_in_proj(jnp.pad(x_t, ((0, 0), (0, 0), (0, self.action_dim - a)))) + token_bias
t_emb = posemb_sincos(time, self.action_in_proj.out_features, min_period=4e-3, max_period=4.0)
t_emb = nnx.swish(self.time_mlp_out(nnx.swish(self.time_mlp_in(t_emb)))) # adaRMS condition
suffix_mask = jnp.ones((b, k), jnp.bool_)
suffix_ar = jnp.array([True] + [False] * (k - 1)) # prefix never attends to the suffix
mask = jnp.concatenate(
[einops.repeat(prefix_mask, "b p -> b s p", s=k), make_attn_mask(suffix_mask, suffix_ar)], axis=-1)
positions = jnp.sum(prefix_mask, axis=-1)[:, None] + jnp.cumsum(suffix_mask, axis=-1) - 1
(_, out), _ = self.PaliGemma.llm(
[None, tokens], mask=mask, positions=positions, kv_cache=kv_cache, adarms_cond=[None, t_emb])
return self.action_out_proj(out)[..., :a]
def timing_interval(self, cond, delta, knots, horizon, cp):
cfg = self.cfg
probes = jnp.broadcast_to(jnp.linspace(0, 1, cfg.timing_probe_count), (cond.shape[0], cfg.timing_probe_count))
values = jax.lax.stop_gradient(bs.evaluate(probes, knots, cp, cfg.degree))
motion, width = values[..., :6], self._gripper01(values)
feats = jnp.concatenate(
[cond, jnp.log(horizon), jax.lax.stop_gradient(delta), motion.reshape(cond.shape[0], -1), width.reshape(cond.shape[0], -1)], -1
)
lower = jnp.minimum(horizon, cfg.min_query_interval)
upper = jnp.minimum(horizon, cfg.max_query_interval)
return lower + (upper - lower) * jax.nn.sigmoid(self.freq_head(feats))
# ---- OpenPI API -----------------------------------------------------------
@override
def compute_loss(self, rng, observation, actions, *, train=False):
cfg = self.cfg
pre_rng, noise_rng, time_rng, ode_rng = jax.random.split(rng, 4)
observation = _model.preprocess_observation(pre_rng, observation, train=train)
acts = actions[..., : cfg.plugin_action_dim].astype(jnp.float32)
b = acts.shape[0]
times = self._times(b)
horizon = self._horizon(b)
if cfg.stage == "timing":
cond, obs_ctx = jax.lax.stop_gradient(self.encode(observation))
knots, delta = jax.lax.stop_gradient(self.predict_knots(cond, horizon))
cp = jax.lax.stop_gradient(
self.sample_control_points(ode_rng, cond, knots, horizon, obs_ctx, cfg.timing_num_ode_steps)
)
interval = self.timing_interval(cond, delta, knots, horizon, cp)
target = labels.timing_target(actions, cfg)
return jnp.square(jnp.log(interval) - jnp.log(target))[:, 0]
cond, obs_ctx = self.encode(observation)
knots, delta = self.predict_knots(cond, horizon)
target_delta = labels.knot_span_target(actions, cfg)
target_cp, _ = bs.fit_control_points(
acts, jax.lax.stop_gradient(knots), times / times[:, -1:], cfg.num_control_points, cfg.degree
)
target_cp = jax.lax.stop_gradient(target_cp)
if self._expert: # pi0.5's flow matching (Pi0.compute_loss): x_t = t*noise + (1-t)*data, u = noise - data
t = jax.random.beta(time_rng, 1.5, 1, (b,)) * 0.999 + 0.001
noise = jax.random.normal(noise_rng, target_cp.shape)
x_t = t[:, None, None] * noise + (1 - t[:, None, None]) * target_cp
v = self._expert_velocity(x_t, t, self._cp_token_bias(knots, horizon), obs_ctx)
flow = jnp.mean(jnp.square(v - (noise - target_cp)), axis=(1, 2))
else:
memory, mask = obs_ctx
t = jax.random.uniform(time_rng, (b,))
x0 = jax.random.normal(noise_rng, target_cp.shape)
x_t = t[:, None, None] * target_cp + (1 - t[:, None, None]) * x0
v = self.flow_policy(x_t, t, self._context(cond, knots, horizon), memory, mask)
flow = jnp.mean(jnp.square(v - (target_cp - x0)), axis=(1, 2))
# KL(target || pred), batchmean -> per-sample sum.
kl = jnp.sum(target_delta * (jnp.log(target_delta) - jnp.log(jnp.maximum(delta, jnp.finfo(delta.dtype).tiny))), -1)
return cfg.flow_loss_coeff * flow + cfg.knot_loss_coeff * kl
@override
def sample_actions(self, rng, observation, *, num_steps=None, noise=None):
"""Dense actions at the control rate (B, action_horizon, action_dim), padded like OpenPI."""
cfg = self.cfg
observation = _model.preprocess_observation(None, observation, train=False)
cond, obs_ctx = self.encode(observation)
b = cond.shape[0]
horizon = self._horizon(b)
knots, _ = self.predict_knots(cond, horizon)
cp = self.sample_control_points(rng, cond, knots, horizon, obs_ctx, num_steps or cfg.num_ode_steps)
times = self._times(b)
dense = bs.evaluate(times / times[:, -1:], knots, cp, cfg.degree)
return jnp.pad(dense, ((0, 0), (0, 0), (0, self.action_dim - cfg.plugin_action_dim)))
def sample_spline(self, rng, observation, num_steps=None):
"""Control points, knots, dense commands, the adaptive query interval and how many commands to execute.
execute_steps = ceil(query_interval / control_dt), extended so that a gripper switch predicted anywhere
in the chunk is executed in this prefix (a switch left after the prefix can be re-predicted after the
prefix forever when the arm stands still and the next observation looks the same), capped at the
chunk length. The client executes dense_actions[:, :execute_steps] and then queries again.
"""
cfg = self.cfg
observation = _model.preprocess_observation(None, observation, train=False)
cond, obs_ctx = self.encode(observation)
b = cond.shape[0]
horizon = self._horizon(b)
knots, delta = self.predict_knots(cond, horizon)
cp = self.sample_control_points(rng, cond, knots, horizon, obs_ctx, num_steps or cfg.num_ode_steps)
interval = self.timing_interval(cond, delta, knots, horizon, cp)
times = self._times(b)
dense = bs.evaluate(times / times[:, -1:], knots, cp, cfg.degree)
steps = jnp.ceil(interval[:, 0] / cfg.control_dt - 1e-4).astype(jnp.int32)
state = labels.gripper_open_state(dense[..., 6])[..., 0] >= cfg.gripper_threshold
switch = state[:, 1:] != state[:, :-1] # switch[j]: state changes at command j + 1
first = jnp.argmax(switch, axis=-1) + 1
steps = jnp.where(jnp.any(switch, -1), jnp.maximum(steps, first + 1), steps)
steps = jnp.clip(steps, 1, self.action_horizon)
return dict(control_points=cp, knots=knots, query_interval=interval, execute_steps=steps,
dense_actions=jnp.pad(dense, ((0, 0), (0, 0), (0, self.action_dim - cfg.plugin_action_dim))))