ACT: Action Chunking with Transformers
8 min read
Predict coherent sequences of joint targets, then blend overlapping plans into smooth, reactive control for delicate bimanual tasks.
the key shift
Predict a short plan, not one twitch
Behavior cloning compounds small errors across a long trajectory. ACT predicts the next k joint targets together, reducing the effective horizon from T decisions to T/k chunks.
Single-step policy
π(aₜ | sₜ)
Every prediction becomes another place where error can push the robot away from the demonstration distribution.
Action-chunk policy
π(aₜ:ₜ₊ₖ | sₜ)
A chunk can encode a coherent subskill, such as grasp, lift, and handover, including pauses that a Markovian one-step policy can misread.
Important nuance: the deployed policy is still queried at every control step. Overlapping chunks are combined with temporal ensembling, so the robot remains reactive.
architecture
A CVAE teaches a Transformer to generate action chunks
During training, the future demonstration supplies a latent style variable. The policy combines that style with current images and joint positions, then decodes the whole action sequence in parallel.
Infer style
CVAE encoder
μ, σ → z = μ + σ ⊙ ε
A BERT-like encoder compresses the action sequence into a latent style. Images are omitted from this training-only path for speed.
Read the scene
Four camera views
ResNet-18 per view, flattened with 2D position embeddings
qₜ + z
State and style
Transformer encoder
Fuses 1,200 visual tokens with projected joint and style tokens into a 1,202 × 512 context sequence.
Decode the chunk
Transformer decoder
k fixed positional queries cross-attend to the encoded observation.
Absolute targets for 6 joints plus one gripper on each arm
cvae
Learn many human styles, deploy one stable style
The latent z explains why equally valid demonstrations can move differently from the same-looking state. It is available from the future only while training.
Training
Reconstruction preserves the demonstrated chunk. KL regularization makes the style space smooth and keeps it near a unit Gaussian.
Inference
Zero is the mean of the training prior, so the deployed policy chooses a deterministic default behavior instead of sampling a new style.
closed-loop execution
Overlap the plans, then agree on this step
A fresh chunk is predicted at every timestep. Several chunks therefore propose an action for the same physical time, and ACT blends those proposals rather than switching plans abruptly.
The highlighted column contains three predictions for action aₜ, all made from different observation times.
The paper uses exponential weights wᵢ = exp(−m · i). Unlike ordinary smoothing across neighboring timesteps, this averages predictions for the same timestep, avoiding temporal bias.
evidence
The gains come from all three pieces
The paper isolates chunking, the CVAE objective, and temporal ensembling, then evaluates ACT on six contact-rich real-world tasks.
1% → 44%
action chunking
Average simulated success rose from k = 1 to k = 100.
2% → 35.3%
CVAE on human data
Success without versus with the CVAE objective.
+3.3%
temporal ensemble
Best ACT success with overlapping predictions combined.
Final real-world task success
25 evaluations per task in the paper
takeaways
The useful intuition
Chunk, but keep looking
Large chunks shorten the learning horizon. Replanning every step and ensembling overlaps preserve visual feedback and smoothness.
Model human variation
The CVAE prevents different valid demonstration styles from collapsing into a hesitant average. At deployment, z = 0 keeps behavior repeatable.
Know the boundary
Original ACT is trained from scratch per task, has no language input, and still struggles when perception or hardware precision is the bottleneck.
In one sentence: ACT learns a coherent distribution over future joint targets, then repeatedly blends overlapping action chunks so a long, delicate task becomes shorter to learn without becoming open-loop at execution time.
The mental model
ACT treats robot control as receding-horizon sequence prediction. Instead of deciding one motor target at a time, it predicts a whole action chunk from four camera views and the robot's current joint positions. This shortens the effective learning horizon and lets a Transformer represent coordinated subskills such as grasping, lifting, and handing an object from one arm to the other.
The CVAE's latent variable absorbs variation in human demonstrations during training. At inference, the encoder is removed and the latent is fixed to zero. The policy still runs at every timestep, and temporal ensembling combines all overlapping predictions for the current action. The result is a useful balance: plan in chunks, observe continuously, and execute smoothly.
Sources