Spaces:
Running on Zero
Running on Zero
File size: 11,394 Bytes
23a59ea | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 | # curiosity.py
"""
MPC-based curiosity exploration policy for active data collection.
Uses CEM (Cross-Entropy Method) to plan action sequences that maximize a
pluggable per-candidate "curiosity score", computed from dynamics-model
predictions. The scorer is injected by the caller (collect_data.py), so the
CEM plumbing here is scorer-agnostic; the published curiosity-driven
collection uses the motion-normalized u_r_norm predictor (see uncertainty.py).
"""
from typing import Callable, Dict, Any, Optional
import torch
from model import Dynamics
from train_dynamics import sample_one_timestep_packed
from uncertainty import sample_predictions_for_actions
# A scorer takes (predictions_KN, z_prev_K) and returns (K,) scalar scores.
# predictions_KN: (K, N, Sz, Dz) float
# z_prev_K: (K, Sz, Dz) float — last observed latent per candidate
# (used for motion normalization)
ScoreFn = Callable[[torch.Tensor, torch.Tensor], torch.Tensor]
@torch.no_grad()
def curiosity_mpc_action(
dyn: Dynamics,
*,
past_packed: torch.Tensor, # (1, t, Sz, Dz)
past_actions: torch.Tensor, # (1, t, A) — actions aligned to past frames
scorer: ScoreFn,
k_max: int,
sched: Dict[str, Any],
act_mask: Optional[torch.Tensor] = None, # (A,) active action dimensions
lang_emb: Optional[torch.Tensor] = None, # (1, lang_dim)
tau_ctx: float = 0.1,
n_candidates: int = 256,
horizon: int = 1,
n_samples: int = 4,
n_elite: int = 32,
n_cem_iters: int = 4,
cem_init_std: float = 1.0,
cem_min_std: float = 0.05,
action_dim: int = 16,
device: Optional[torch.device] = None,
use_kv_cache: bool = False,
) -> Dict[str, torch.Tensor]:
"""
Select an action (or action sequence) that maximizes the given per-candidate
score via CEM.
Returns dict with:
"action": (A,) — first action of the chosen plan in [-1, 1]
"action_sequence": (H, A) — full plan (only present for horizon > 1)
"uncertainty": scalar — score of the chosen plan
"mean_uncertainty": scalar — mean score across the last CEM iter's candidates
"max_uncertainty": scalar — max score across the last CEM iter's candidates
("uncertainty" is the generic name for the injected scorer's signal.)
"""
if device is None:
device = past_packed.device
dtype = next(dyn.parameters()).dtype
past_packed = past_packed.to(dtype)
past_actions = past_actions.to(dtype)
A = action_dim
if act_mask is not None:
if act_mask.dim() == 1:
active_mask = act_mask # (A,)
else:
active_mask = act_mask[0, 0] # (A,) from (B, T, A)
else:
active_mask = torch.ones(A, device=device)
if horizon == 1:
return _cem_single_step(
dyn=dyn,
past_packed=past_packed,
past_actions=past_actions,
scorer=scorer,
k_max=k_max, sched=sched,
act_mask=act_mask, lang_emb=lang_emb, tau_ctx=tau_ctx,
n_candidates=n_candidates, n_samples=n_samples,
n_elite=n_elite, n_cem_iters=n_cem_iters,
init_std=cem_init_std, min_std=cem_min_std,
action_dim=A, active_mask=active_mask, device=device,
)
return _cem_multi_step(
dyn=dyn,
past_packed=past_packed,
past_actions=past_actions,
scorer=scorer,
k_max=k_max, sched=sched,
act_mask=act_mask, lang_emb=lang_emb, tau_ctx=tau_ctx,
n_candidates=n_candidates, horizon=horizon, n_samples=n_samples,
n_elite=n_elite, n_cem_iters=n_cem_iters,
init_std=cem_init_std, min_std=cem_min_std,
action_dim=A, active_mask=active_mask, device=device,
use_kv_cache=use_kv_cache,
)
@torch.no_grad()
def _cem_single_step(
dyn, past_packed, past_actions, scorer,
k_max, sched, act_mask, lang_emb, tau_ctx,
n_candidates, n_samples, n_elite, n_cem_iters,
init_std, min_std,
action_dim, active_mask, device,
) -> Dict[str, torch.Tensor]:
# Gaussian CEM (same scheme as plan_cem.py): sample (mu + std * noise).clamp(-1, 1),
# refit mean/std on elites, floor std at min_std.
mu = torch.zeros(action_dim, device=device)
sigma = torch.full((action_dim,), float(init_std), device=device)
best_action = None
best_score = torch.tensor(-float("inf"), device=device)
# Motion reference: last observed latent, broadcast to K at score time.
z_last = past_packed[0, -1].float() # (Sz, Dz)
for cem_iter in range(n_cem_iters):
noise = torch.randn(n_candidates, action_dim, device=device)
candidate_acts = (mu + sigma * noise).clamp(-1, 1)
candidate_acts = candidate_acts * active_mask.unsqueeze(0)
past_acts_K = past_actions.expand(n_candidates, -1, -1) # (K, t, A)
new_act = candidate_acts.unsqueeze(1) # (K, 1, A)
full_actions = torch.cat([past_acts_K, new_act], dim=1) # (K, t+1, A)
predictions = sample_predictions_for_actions(
dyn,
past_packed=past_packed,
candidate_actions=full_actions,
k_max=k_max, sched=sched,
act_mask=act_mask, tau_ctx=tau_ctx,
lang_emb=lang_emb, n_samples=n_samples,
) # (K, N, Sz, Dz)
z_prev_K = z_last.unsqueeze(0).expand(n_candidates, -1, -1) # (K, Sz, Dz)
scores = scorer(predictions, z_prev_K) # (K,)
max_idx = scores.argmax()
if scores[max_idx] > best_score:
best_score = scores[max_idx]
best_action = candidate_acts[max_idx]
_, elite_idx = scores.topk(n_elite)
elite_acts = candidate_acts[elite_idx]
mu = elite_acts.mean(dim=0)
sigma = elite_acts.std(dim=0).clamp_min(float(min_std))
return {
"action": best_action,
"uncertainty": best_score,
"mean_uncertainty": scores.mean(),
"max_uncertainty": scores.max(),
}
@torch.no_grad()
def _cem_multi_step(
dyn, past_packed, past_actions, scorer,
k_max, sched, act_mask, lang_emb, tau_ctx,
n_candidates, horizon, n_samples, n_elite, n_cem_iters,
init_std, min_std,
action_dim, active_mask, device,
use_kv_cache: bool = False,
) -> Dict[str, torch.Tensor]:
t = past_packed.shape[1]
# Gaussian CEM (same scheme as plan_cem.py) with per-timestep (mu, std):
# sample (mu + std * noise).clamp(-1, 1), refit on elites, floor std at min_std.
mu = torch.zeros(horizon, action_dim, device=device)
sigma = torch.full((horizon, action_dim), float(init_std), device=device)
best_action_seq = None
best_total = torch.tensor(-float("inf"), device=device)
for cem_iter in range(n_cem_iters):
noise = torch.randn(n_candidates, horizon, action_dim, device=device)
cand_seqs = (mu.unsqueeze(0) + sigma.unsqueeze(0) * noise).clamp(-1, 1)
cand_seqs = cand_seqs * active_mask.unsqueeze(0).unsqueeze(0)
totals = _rollout_scores(
dyn=dyn,
past_packed=past_packed,
past_actions=past_actions,
candidate_seqs=cand_seqs,
scorer=scorer,
k_max=k_max, sched=sched,
act_mask=act_mask, lang_emb=lang_emb, tau_ctx=tau_ctx,
n_samples=n_samples,
max_ctx=t,
use_kv_cache=use_kv_cache,
) # (K,)
max_idx = totals.argmax()
if totals[max_idx] > best_total:
best_total = totals[max_idx]
best_action_seq = cand_seqs[max_idx]
_, elite_idx = totals.topk(n_elite)
elite_seqs = cand_seqs[elite_idx]
mu = elite_seqs.mean(dim=0)
sigma = elite_seqs.std(dim=0).clamp_min(float(min_std))
return {
"action": best_action_seq[0],
"action_sequence": best_action_seq,
"uncertainty": best_total,
"mean_uncertainty": totals.mean(),
"max_uncertainty": totals.max(),
}
@torch.no_grad()
def _rollout_scores(
dyn, past_packed, past_actions, candidate_seqs, scorer,
k_max, sched, act_mask, lang_emb, tau_ctx, n_samples,
max_ctx: int = 0,
use_kv_cache: bool = False,
) -> torch.Tensor:
"""
Roll out K candidate action sequences through the dynamics model,
scoring per horizon step and returning an aggregated score per candidate.
Aggregation: 0.5 * mean_per_step + 0.5 * max_per_step (balances average
vs. peak score along the plan).
"""
K, H, A = candidate_seqs.shape
t = past_packed.shape[1]
Sz, Dz = past_packed.shape[2], past_packed.shape[3]
past_K = past_packed.expand(K, -1, -1, -1) # (K, t, Sz, Dz)
past_acts_K = past_actions.expand(K, -1, -1) # (K, t, A)
lang_K = None if lang_emb is None else lang_emb.expand(K, -1)
z_history = [past_K[:, i] for i in range(t)]
act_history = [past_acts_K[:, i] for i in range(t)]
sum_score = torch.zeros(K, device=past_packed.device)
max_score = torch.zeros(K, device=past_packed.device)
for h in range(H):
act_h = candidate_seqs[:, h] # (K, A)
if max_ctx > 0 and len(z_history) > max_ctx:
z_win = z_history[-max_ctx:]
act_win = act_history[-max_ctx:]
else:
z_win = z_history
act_win = act_history
ctx_len = len(z_win)
z_seq = torch.stack(z_win, dim=1) # (K, ctx_len, Sz, Dz)
acts_seq = torch.stack(act_win + [act_h], dim=1) # (K, ctx_len+1, A)
# Run N samples per candidate in one batch (K*N).
KN = K * n_samples
z_KN = z_seq.unsqueeze(1).expand(-1, n_samples, -1, -1, -1).reshape(KN, ctx_len, Sz, Dz)
acts_KN = acts_seq.unsqueeze(1).expand(-1, n_samples, -1, -1).reshape(KN, ctx_len + 1, A)
lang_KN = None if lang_K is None else lang_K.unsqueeze(1).expand(-1, n_samples, -1).reshape(KN, -1)
z_next = sample_one_timestep_packed(
dyn,
past_packed=z_KN,
k_max=k_max, sched=sched,
actions=acts_KN, act_mask=act_mask, tau_ctx=tau_ctx,
lang_emb=lang_KN,
use_kv_cache=use_kv_cache,
) # (K*N, Sz, Dz)
predictions = z_next.float().reshape(K, n_samples, Sz, Dz)
z_prev_K = z_win[-1].float() # (K, Sz, Dz)
score_h = scorer(predictions, z_prev_K) # (K,)
sum_score = sum_score + score_h
max_score = torch.max(max_score, score_h)
# Mean prediction drives AR continuation.
z_mean = predictions.mean(dim=1).to(past_packed.dtype) # (K, Sz, Dz)
z_history.append(z_mean)
act_history.append(act_h)
return 0.5 * (sum_score / max(H, 1)) + 0.5 * max_score
|