inputs: x: (B, Ta, Da) time: (B,) floating point in [0,1) flow matching time cond: dict with key state/rg

inputs: action: (B, Ta, Da) action chunk time: (B,) or float within [0,1), flow time cond: dict with key

inputs: x: (B, Ta, Da) time: (B,) floating point in [0,1) flow matching time cond: dict with key state/rg

Generate actions from observations and noise using the actor network. Args: cond (Dict[str, Tensor]): Dictionary containing the s

x: [batch, seq, embed_dim]

cond: dict with key state/rgb; more recent obs at the end state: (B, To, Do) action: (B, Ta, Da)

cond: dict with key state/rgb; more recent obs at the end state: (B, To, Do) rgb: (B, To, C, H, W) no_augment

cond: (B, T, cond_dim) output: (B, T, output_dim)

inputs: cond: dict, contatinin... 'state': obs. observation in robotics. torch.Tensor(batchsize, cond_steps, obs_

Returns all actions from all trajectories, concatenated on dim 0 (time).

Returns all actions from all trajectories, concatenated on dim 0 (time).

Get a property from each sub-environment. Parameters ---------- name : string Name of the property to be get from

inputs: x_chain: torch.Tensor of shape `[batchsize, self.inference_steps+1, self.horizon_steps, self.act_dim]`

for training and validation on fixed dataset

for training and validation on fixed dataset

for training and validation on fixed dataset

Calculates the difference between the current quaternion and the desired quaternion. See Siciliano textbook page 140 Eq 3.91 :param curr

Returns the length of the idx-th trajectory.

for testing purpose

for testing purpose

for testing purpose

ViT weight initialization, original timm impl (for reproducibility)

Compute an unbiased estimate to the log-likelihood in bits/dim. Args: cond: dict with key state/rgb; more recent obs at the end

REINFORCE loss. Not used right now. Args: cond: dict with key state/rgb; more recent obs at the end stat

PPO loss obs: dict with key state/rgb; more recent obs at the end state: (B, To, Do) rgb: (B, To, C, H, W)

PPO loss obs: dict with key state/rgb; more recent obs at the end state: (B, To, Do) samples: (B, Ta, Da)

Compute the combined loss for flow-matching (d=0) and self-consistency (d>0). Args: x1: (B, Ta, Da) - Real action trajec

PPO loss obs: dict with key state/rgb; more recent obs at the end "state": (B, To, Do) "rgb": (B, To, C, H, W

PPO loss obs: dict with key state/rgb; more recent obs at the end state: (B, To, Do) rgb: (B, To, C, H, W)

GRPO loss Inputs: obs: dict with key state/rgb; more recent obs at the end state: (B, To, Do)

PPO loss obs: dict with key state/rgb; more recent obs at the end state: (B, To, Do) rgb: (B, To, C, H, W)

PPO loss obs: dict with key state/rgb; more recent obs at the end state: (B, To, Do) rgb: (B, To, C, H, W)

PPO loss obs: dict with key state/rgb; more recent obs at the end state: (B, To, Do) rgb: (B, To, C, H, W)

no squashing

Compute the actor loss combining behavior cloning, Q-value, and distillation losses, with debug information. Args: obs (Dict[str,